Automated Detection and Classification of Nipple Damage in Lactation Care

Automated Detection and Classification of Nipple Damage in Lactation Care | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Automated Detection and Classification of Nipple Damage in Lactation Care Jessica de Souza, Kelly Pereira Coca, Bárbara Tideman Sartorio Camargo, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6348223/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Lactation-related nipple damage is a prevalent issue among breastfeeding mothers, often leading to early breastfeeding cessation due to pain and misdiagnosis. Accurate and timely classification of nipple damage is critical for effective treatment, yet current methods rely on subjective clinical assessments, resulting in variability and inefficiency. This study addresses these challenges by developing a Deep Learning (DL) system for the automated detection and classification of nipple damage. Using a dataset of 1,090 images from clinical trials developed in São Paulo, Brazil, we implemented a Resnet50 convolutional neural network (CNN) to perform two tasks: (1) binary classification to differentiate between intact nipples and those with damages and (2) multiclass classification to identify four types of damage (closed wound, crust, erosion, and fissure) based on the instrument for classifying nipple and areola complex lesions. Data augmentation techniques were applied to upsample the dataset to 8,720 images. The binary classification model achieved an average area under the receiver operating characteristics curve (AUROC) of 0.99 and a recall of 95.90%, demonstrating high accuracy in detecting nipple damage. The multiclass model achieved AUROC values ranging from 0.89 to 0.99 in nipple damage classification, with the highest performance observed for closed wounds (AUROC = 0.98) and erosion (AUROC = 0.99). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed the model’s focus on damaged areas, which aligned closely with clinical assessments. Our findings highlight the potential of DL to improve lactation care by enabling accurate, automated nipple damage classification, particularly in settings with limited access to lactation specialists. This study represents a significant step toward leveraging technology to address challenges in lactation care and improve outcomes for breastfeeding mothers. Health sciences/Health care Health sciences/Health care/Diagnosis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Lactation-related nipple damage is a common condition affecting breastfeeding mothers, with studies estimating that 13–92% experience some degree of this complication within the first three weeks postpartum, often leading to early breastfeeding cessation [ 1 – 4 ]. Causes of nipple damage include improper latching and positioning of the infant onto the breast, infections, infant anatomic anomaly (i.e., tongue-tie, high palate), and nipple anatomy [ 4 – 7 ]. Once a professional identifies the nipple damage, assesses its cause, and proposes a treatment, it can take several days to heal based on its initial size and depth [ 4 , 7 ]. Proper nipple damage assessments by health professionals are crucial for effective and timely treatments, providing patients with adequate healing time that fits the maintenance of breastfeeding [ 4 ]. Misdiagnosis of nipple damage and lack of consensus on treatment between health professionals can prolong treatment times, burden the patient with physical and mental pain, cause infant distress, and put the continuity of breastfeeding at risk [ 8 , 9 ]. Current methods for nipple damage classification typically involve detailed clinical examination, visual inspection with accessories for better visualization (i.e., direct light, magnifying glass), measuring tools and scales, and digital image acquisition for clinical discussion and monitoring damage progression [ 4 , 10 ]. Classifying nipple damage by type is a time-consuming and subjective process, often leading to variability in assessments by different health providers [ 11 ]. Recent advancements in nipple damage research have brought standardization of damage assessment, such as the “Classification Instrument for Nipple-Areolar Lesions” (in its original naming in Brazilian Portuguese: Instrumento de Classificação das Lesões Mamilo-Areolares) ( ILMA ) by Cervellini et al. 2022 [ 12 ] and the instrument “Seven Signs of Nipple Trauma Associated with Breastfeeding” from Nakamura et al. 2018 [ 13 ]. These studies emphasize the importance of proper terminology of damage under a dermatological approach, as nipple damage is frequently misclassified under the broad umbrella of 'fissures' without accounting for distinct lesion characteristics [ 2 ]. This happens due to a lack of consensus in the literature on nipple injury nomenclature [ 10 ] and reduced specialized lactation training, as the number of board-certified IBCLCs is still rising [ 14 ]. Classifying the damage by type and damage degree, as well as identifying the cause, can contribute to a more specific treatment and bring more satisfactory results for patients who suffer from nipple pain. The widespread adoption of Deep Learning (DL) in women’s health applications has benefited clinical decision-making in several fields and enhanced patient care for better accuracy when diagnosing conditions, ranging from cervical and breast cancer exam assessments [ 15 – 18 ] to maternal and fetal well-being predictions [ 19 , 20 ]. Only recently, DL was adopted in lactation care to identify breastfeeding-related conditions. Firstly, Convolutional Neural Networks (CNNs) were applied to breast and nipple images to detect different lactation complications, such as abscess, mastitis, nipple blebs, dermatosis, engorgement, and nipple damage [ 21 ]. De Souza et al. used a dataset of 1078 images and applied data augmentation techniques, achieving an area under the receiver operating characteristics curve (AUROC) of 0.93 for the best model. This demonstrated the potential of algorithms to assist healthcare providers in identifying complications and their applicability for at-home tele-triaging in the first few weeks postpartum [ 21 , 22 ]. Another example of DL in lactation care is the work from Nakamura et al., which employed a nipple damage evaluation system using their previous measurement tool, “Seven Signs of Nipple Trauma Associated with Breastfeeding,” which used a dataset of 753 images and achieved an average AUROC above 0.7 to classify the classes erythema, swelling, blistering, scabbing, fissure, purpura, peeling and none [ 23 ]. These previous studies highlighted the potential of using new technologies to address critical inequities in lactation care access and improve outcomes for mothers where there is a lack of professional availability and expertise. However, these studies also uncovered several challenges in classifying breastfeeding complications and nipple damage, mainly related to limited data, such as small image samples for using more robust DL algorithms, lack of racial and geographic diversity due to data collection limitations, which may not generalize well in diverse ethnic groups, and imbalanced samples in categories that are harder to classify between, that is also a problem in the real-world [ 21 , 23 ]. This study addresses previous work gaps by developing an automated system for detecting and classifying nipple damage/lesions using DL techniques. We build on previous research by using a dataset from clinical trials in São Paulo, Brazil, and online gathered images from a previous study. This study focuses on applying a ResNet50 Convolutional Neural Network (CNN) to perform binary and multiclass classification tasks, helping clinicians differentiate between healthy nipples and various types of nipple damage. This work also introduces a novel approach by incorporating a racially diverse dataset to ensure the model's broader applicability, especially in Latin-American contexts. Our main contributions are listed as follows: We present the feasibility of using DL models to detect and classify nipple damage using image classification techniques while providing a detailed methodology for application reproducibility. We evaluate the model’s ability to generalize to unseen data by applying stratified 10-fold cross-validation, reducing bias in our metrics, and maximizing data use. We propose end-to-end image preprocessing, data augmentation techniques, and evaluation methods to overcome challenges when handling a limited dataset. We discuss the implications of automated damage classification systems in real-world settings and provide insights into future designs for adequately applying the system in clinical settings. Results Overview We evaluated the ability of the Resnet50 model to classify nipple images across two tasks. The first task was the binary classification to assess whether the images had damage. The second task was a multi-class classification of the type of damage across four classes. The dataset used had 7936 images in the training and validation set and 784 images on the independent test set for the binary classification task. We used the same dataset for the second task but removed the intact images, totaling 6888 images for the training and validation set and 688 images for the independent test set. As part of evaluating both tasks, we implemented stratified 10-fold cross-validation to test the model’s generalization ability within the imbalanced dataset. There was no hyperparameter tuning between each fold, and all tasks used the same optimizer, learning rate, weight decay, and batch size. We share the model’s performance in the binary and multiclass classifications in the following subsections. We present the following performance metrics, divided into overall average and per fold: training and validation accuracy, false negatives (FN), false positives (FP), precision, recall, and F1-score. The macro-averaged metrics are reported after concatenating the inferences in the test set from all folds. In our model, we defined equal importance in detecting each class, even with an imbalanced dataset, therefore, a macro average was appropriate. Additionally, we present figures with additional metrics for the two tasks: The Area Under the Receiver Operating Characteristic Curve (AUROC), the Receiver Operating Characteristic (ROC) curve, and the Kernel Density Estimate (KDE) plot and its respective boxplot with the distribution of predicted probabilities. To enhance our understanding of the model outputs, we employ an aggregated confusion matrix that consolidates the predictions across all 10 iterations applied to the data set. We achieved this aggregation by taking the median predicted class for each instance over the multiple folds, synthesizing a singular prediction representing the agreement of the model’s behavior across the independent test set. Evaluation of Model’s Performance on Binary Classification Table 1 shows the results for 10-fold cross-validation, demonstrating the model’s consistent performance across the iterations. In Fig. 1 , we present several performance metrics from the model: Fig. 1 a shows the aggregated confusion matrix for the Resnet50 model. Figure 1 b displays the AUROC curve for each class, where the damage class has an AUROC of 0.99 (95% CI, 0.99 to 1.00). Figure 1 c shows the KDE plot and the boxplot with the distribution of predicted probabilities for both classes. We can observe minimal overlap between the classes in the distribution, where the probability below 0.50 corresponds to the class with lesion, and the probability above 0.5 corresponds to the intact class. Table 1 Results of 10-fold cross-validation for the augmented data set on the binary class damage detection task using the Resnet50 model. 10-fold iterations Training Accuracy Validation Accuracy Precision Recall F1-score Iteration 1 0.995 0.981 0.921 0.973 0.945 Iteration 2 0.996 0.979 0.923 0.983 0.951 Iteration 3 0.996 0.982 0.917 0.947 0.931 Iteration 4 0.995 0.984 0.921 0.973 0.945 Iteration 5 0.995 0.977 0.922 0.978 0.948 Iteration 6 0.996 0.982 0.921 0.973 0.945 Iteration 7 0.996 0.982 0.917 0.973 0.943 Iteration 8 0.995 0.982 0.923 0.963 0.942 Iteration 9 0.996 0.984 0.917 0.973 0.943 Iteration 10 0.996 0.987 0.928 0.968 0.947 Based on the aggregated confusion matrix, out of the 784 images used in the testing set, the model could correctly classify 762 images, giving a recall of 95.90%, a precision of 91.72%, and an F1-score of 93.67%. The total images correctly classified by category are as follows: damage is the best class (671/688; accuracy = 98%, F1-score = 0.984), followed by intact (91/96; accuracy = 95%, F1-score = 0.890). Table 2 summarizes the model’s performance per class on the augmented test set. Table 2 Summary of the detection results per class: accuracy, precision, recall, F1-score, and support (ie, number of samples per class) using the Resnet50 architecture. Class Name FP (N) Precision Recall F 1 -score Support Damage 17 0.992 0.975 0.984 688 Intact 5 0.843 0.943 0.890 96 Evaluation of Model’s Performance on Multiclass Classification Table 3 shows the results for 10-fold cross-validation on the multiclass classification task, highlighting the model’s consistent performance across the iterations. Figure 2 a shows the aggregated confusion matrix from the Resnet50 model. Figure 2 b shows the AUROC curve for each class, with the following values: closed wound (AUROC = 0.98, 95% CI, 0.98 to 0.98), crust (AUROC = 0.90, 95% CI, 0.90 to 0.91), erosion (AUROC = 0.99, 95% CI, 0.99 to 0.99), and fissure (AUROC = 0.89, 95% CI, 0.89 to 0.90). Figure 2 c shows the KDE plot and the boxplot illustrating the distribution of predicted probabilities for each class. To estimate which class was predicted in the test set, we used a one-hot encoder [ 24 ] from the sklearn library [ 25 ], which evaluated the output probabilities for each class and binarized the highest probability as one and the lowest probabilities as zero. Table 3 Results of 10-fold cross-validation for the augmented data set on the multi-class damage classification task using the Resnet50 model. 10-fold iterations Training Accuracy Validation Accuracy Precision Recall F1-score Iteration 1 0.913 0.863 0.833 0.844 0.836 Iteration 2 0.916 0.804 0.836 0.839 0.835 Iteration 3 0.912 0.805 0.837 0.840 0.836 Iteration 4 0.915 0.840 0.840 0.842 0.839 Iteration 5 0.914 0.839 0.820 0.841 0.826 Iteration 6 0.910 0.839 0.848 0.854 0.850 Iteration 7 0.913 0.828 0.830 0.840 0.833 Iteration 8 0.913 0.843 0.842 0.841 0.840 Iteration 9 0.915 0.799 0.832 0.843 0.833 Iteration 10 0.919 0.828 0.848 0.837 0.840 Based on the aggregated confusion matrix, out of the 688 images used in the testing set, the model could correctly classify 542 images, giving a macro-average precision of 76.83%, recall of 82.07%, and F1-score of 78.96%. The total images correctly classified by category are as follows: closed wound (88/96; accuracy = 92%), crust (250/320; accuracy = 78%), erosion (65/72; accuracy = 90%), and fissure (139/200; accuracy = 69%). Among the remaining images that were misclassified with a percentage of error higher than 10%, the class crust had 12% (37/320) of images misclassified as fissure, the class fissure had 21% (42/200) of images misclassified as crusts, and the class erosion had 10% (7/72) misclassified as crusts. The best-performing class was erosion (F1-score = 0.829), followed by closed wound (F1-score = 0.809), crust (F1-score = 0.803), and fissure (F1-score = 0.718). Table 4 summarizes the model’s performance per class on the augmented test set. Table 4 Summary of the detection results per class: accuracy, precision, recall, F1-score, and support (ie, number of samples per class) using the Resnet50 architecture. Class Name FN (N) FP (N) Precision Recall F 1 -score Support Closed wound 8 31 0.726 0.914 0.809 96 Crust 70 52 0.830 0.779 0.803 320 Erosion 7 22 0.765 0.907 0.829 72 Fissure 61 41 0.757 0.683 0.718 200 Figure 3 shows image samples for true positives and false negatives the original image accompanied by the image regions responsible for the model’s decision-making process through the GRAD-CAM activation maps. By overlaying the activation map on top of the original image, we can observe how the model highlights the same area corresponding to the damage and region of interest (highlighted in cyan in the original image). Each of the four classes with damages have different patterns in the activation map, where the warmest colors show areas with higher probability in the class of interest. Discussion This study evaluated the potential of DL in identifying nipple damage through image classification using an adaptation of the ILMA tool [ 12 ], which used two independent models to detect a damage through the binary task and classify the type of damage through the multiclass task. Through comprehensive testing on the independent test set, we demonstrate the robust performance and the ability of our model to distinguish between healthy and unhealthy images, as well as identify four types of nipple damage, such as damages without interruption of the skin barrier (closed wound), crusts, erosions, and fissures with high precision, recall, and F1-score. It is essential to note the model's excellent ability in the binary classification task. By identifying whether there are signs of damage in a nipple image evaluation, this model alone can help triage and indicate patients who potentially are facing some disruption and discomfort in their breastfeeding journey. Meanwhile, classifying the damage by type can help health professionals with more accurate diagnoses, allowing targeted treatments for improved healing outcomes. When analyzing the misclassified images in the independent test set, several factors were identified that may have contributed to the model's confusion. These include: (a) Similar visual features between classes, such as overlapping characteristics of different damage types. (b) Variations in nipple-areolar color and anatomy can affect the model's ability to distinguish between classes. (c) Environmental lighting during image acquisition, particularly in images taken with low-light conditions, where shadows on darker areolas resembled crusting in fissure images. (d) Images showing multiple damages simultaneously, such as a crust image with micro-fissures, complicate the classification process, and (e) conditions that precede or evolve into one another, such as erosion healing into a crust over time. These challenges were also mentioned in previous work and are not just technical limitations, as they also happen in real-world settings and can impact professional decision-making [ 21 , 23 ]. To provide further insight into the binary classification task, images with closed wound damages were occasionally misclassified as intact and vice versa, particularly in lighter skin tones and pink areolas. Similarly, confusion between crust and erosion damages arose due to the color variations in scabbing that varied (from red, light yellow to brown), overlapping with the red, fluid-filled appearance of some erosion images. Additionally, some images were taken over several days with intervals of 12 hours, capturing the transition of an erosion damage healing into a crust, which may have further complicated the model's ability to differentiate between the two. Another challenge was the presence of nipple anatomy variation, e.g., inverted and protruding nipples with a granular texture, resembling a "raspberry nipple" [ 26 ], which could be mistaken for vesicles or blistering despite being healthy. While the model may have learned certain features from the same user across different images, variations in lighting and angles likely influenced its performance. Although data augmentation techniques helped mitigate overfitting and increased image variety, some issues persisted. However, these challenges do not significantly impact the model's overall performance, particularly in cases where damages naturally evolve, such as an erosion transitioning to a crust. The ability to classify nipple damage using DL with the ILMA classification tool offers significant potential to standardize and promote the use of proper terminologies, fostering greater adoption among healthcare professionals. To maximize the benefits of this technology, health specialists should receive targeted education on nipple damage classification to ensure accurate diagnosis and streamlined patient treatment [ 2 ]. This algorithm has several practical applications: (1) Professional Training: It can be an educational tool to train healthcare providers on the diverse types of nipple damage, improving their diagnostic skills and confidence. (2) Support in Underserved Areas: In regions with limited access to specialized lactation care, the algorithm can assist general health providers in accurately classifying damages and recommending appropriate treatments. (3) Remote Classification: The tool can enable remote damage classification in healthcare facilities with limited professional availability, ensuring timely and accurate assessments. (4) Patient Triage and Hybrid Care: For lactation providers offering hybrid care models, the algorithm can help prioritize patients based on damage severity and streamline workflows [ 22 ]. (5) Postpartum Monitoring: When integrated with ecological momentary assessment (EMA), the tool can facilitate regular checkups during the first few weeks postpartum, aiding in early damage recognition and prevention [ 27 ]. Some limitations of this work include the limited dataset size in specific classes, such as within the closed wounds group, limiting the model’s ability to separate the categories that have no skin barrier interruption, such as erythema, edema, vesicle, and ecchymosis and lead to the grouping of images into one condition. Another limitation was the lack of clinical standardization for conditions by severity to a more specialized classification. While previous work proposed a severity-based nipple trauma classification (i.e., none, minor, moderate, and severe) [ 23 ], the clinical adoption of such tiers remains limited due to the absence of standardized, validated criteria, such as using the Nipple Trauma Score (NTS) tool proposed by Abou-Dakn et al. [ 28 ]. Measuring lesion severity is a subjective task, leading to the lack of professional agreement on the level of damage severity in real-world scenarios [ 10 , 33 ]. Given these challenges, our model focused on lesion detection and classification rather than severity grading, which is aligned with current clinical practices [ 12 ]. Future works include (1) assessing the dataset images using the NTS tool across and performing a DL evaluation to assess the capacity for automated NTS detection, (2) validating a practical tool for real-time classification across multidisciplinary providers while evaluating model accuracy, professional trust, comfort, and privacy concerns in using technology to assist in their work. And (3), expand the dataset to incorporate higher racial diversity while increasing the distribution of different conditions and including additional variables such as pain levels, persistence of the damage, and treatments applied to follow damage progression. In conclusion, our study introduces an innovative DL model capable of automatically detecting and classifying nipple damage using the ILMA classification tool. The accurate assessment of breast conditions is crucial, as pain and discomfort during breastfeeding represent significant barriers for parents striving to continue breastfeeding their infants [ 29 , 30 ]. We hope this work will open meaningful discussions on how technology can address critical inequities in access to lactation care and support professional education on the topic, helping to standardize the nomenclature of nipple damage while improving treatment outcomes. By supporting healthcare providers in regions with limited professional availability and expertise, this tool has the potential to improve patient outcomes, reduce the distress caused by nipple damage, and ultimately enhance the breastfeeding experience for mothers worldwide. Methods Study Design Ethical Considerations This multi-site study was approved by the Institutional Review Boards of the University of California, San Diego (No. 801,904) and the Federal University of São Paulo (No. 1710/09, No. 1.267.605, and No. 4.726.329). No personally identifiable data from the participants were incorporated into this research. Study Type and Data Sources This study comprises two separate datasets on lactation-related nipple damage. The first dataset had the photographic images obtained through clinical settings, where participants were in the first two days postpartum, between 2011 and 2024, in São Paulo, Brazil. The detailed protocol for participant selection and data collection is described in previous publications [ 31 – 33 ]. The second dataset was gathered in a previous study and used mixed-media data collection (i.e., physical and online clinical resources) [ 21 ]. The second dataset was included in this study to complement the clinical dataset and help with balancing classes for better algorithm generalization. We performed a retrospective analysis of the prospectively collected data before including them in the algorithm for detection and classification. Dataset Characteristics Participants A total of 1219 nipple images were collected to compose the dataset, described in Table 5 . The images were categorized into primary damage type, nipple color, and damage location. Following applying the inclusion criteria, 1,090 images were retained, while 129 images with visual issues that could compromise the study were excluded. The images were categorized into five classes, following the ILMA [ 12 ] classification instrument: Closed wounds, crust, erosion, fissure, and intact. The closed wound class presents Nipple and Areola Damages (NADs) without skin barrier disruption, including erythema, ecchymosis, vesicle, and edema. This class was concatenated due to visual similarities in the conditions and the limited number of images needed to compose separate classes. The crust class shows scabbing of the skin and presents itself after the damage is healing and can have yellow, brown, or dark red tones. The erosion class presents damages with well-defined borders and can contain clear or blood fluid, exposing the skin's dermis or epidermis. The fissure class presents damages with a narrow and linear shape, which may present varied damage depths, with or without the presence of fluids. Finally, the intact class includes images without classification from the ILMA instrument, where images presenting healthy tissue and no damage were classified into this category. The classification of the nipple-areolar color presented three primary colors, where we categorized images in black, brown, and pink pigmentation [ 34 ], with a variation between lighter and darker pigmentation that was not considered since the damages often presented themselves with redness at the nipple surface. No Fitzpatrick Skin Tone classification was performed in the images since no skin surface was evaluated in the study, only nipple-areolar tissue. Inclusion and Exclusion Criteria To be included in the dataset, the images had to meet the following criteria: (1) RGB format (PNG or JPEG); (2) visually present at least 1 of the 5 conditions; (3) have the nipple visible; (4) have no irrelevant objects covering the nipple; (5) have sufficient resolution for clinical evaluation, showing no signs of low quality, such as blurriness (due to motion or focus issues), pixelation, overexposure or underexposure. Images were excluded from the data set if (1) the breast or nipple was from non-lactating female patients; (2) there was no match between the image and the label; (3) the breast or nipple was not visible in the image; and (4) the image had no label describing it. Table 5 Original dataset characteristics. Category Details Dataset Size (MB) 344.27 Image Size (KB) Min: 6.46, Average: 326.12, Max: 1570.42 Image Dimensions (pixels) Min: 77 × 77, Average: 740 × 740, Max: 1773 × 1773 Data Collection Period 2011 to 2024 Data Source Prior Clinical Trials in São Paulo: 847 [ 31 – 33 ] Mixed-media source from previous study: 243 [ 21 ] Total Images Collected 1219 Total Images Excluded 129 (10.58%) Final Dataset Size 1090 (89.42%) Number of Classes 5 Number of Unique Subjects 403 Images per Damage Category Closed wounds: 141 Crust: 389 Erosion: 100 Fissure: 317 Intact: 143 Predominant Nipple Color Black: 57 Brown (light to darker tones): 901 Pink (light to darker tones): 132 Nipple Damage Location Apex: 864 Lateral: 83 Non-present (intact): 143 Data Labeling Protocol Two board-certified nurse practitioners with a PhD in Nursing, specialists in maternal and lactation care, and International Board-Certified Lactation Consultants (IBCLC) with over 10 years of experience, collaboratively classified the images on the dataset to ensure that their classifications had no discrepancies. This consensus-based classification process involved the professionals reviewing each image together and discussing the most prominent damage until they reached an agreement. The damage classification was based on the ILMA instrument [ 12 ]. The entire classification process took approximately six hours. Image Preprocessing Before using the images as inputs for the DL models, the images were manually cropped in a 1:1 ratio to prevent image flattening or warping during resizing and loss of essential features. Most images have the nipple tissue concentrated in the center of the image, thereby focusing the model’s evaluation on the most relevant areas. We followed the guidelines proposed by similar studies on DL for skin disease and damage classification [ 35 – 38 ], which aim to objectively show the area of interest for optimized detection and reduce the risks of poorly triaged images. The image preprocessing pipeline and samples of the augmentations used are shown in Fig. 4 . Before entering the DL pipeline, we applied standard transformations in the data, starting with image resizing. In this paper, we trained, validated, and tested our data set using the Resnet50 Convolutional Neural Network (CNN). This CNN requires the input images to be resized to 224×224 pixels. Therefore, we proceeded with the image resizing according to each model’s requirements. The last transformation step incorporates the normalization of the images, which standardizes pixel intensity across the data set. For better model generalization in this dataset, we calculated the mean and SD of all images in the data set to use in the normalization process instead of using the ImageNet data set pre-trained parameters, following previous study involving breast disease classification [ 21 ]. To increase the variety of the dataset while improving the model’s robustness and reducing the risk of overfitting, we implemented data augmentation techniques to upsample the images on the dataset (Fig. 4 ). We used the following seven data augmentations previously used in data sets involving skin damage [ 36 , 39 ]: brightness, random affine, contrast, horizontal and vertical flip, center zoom, and rotation. Before data augmentation, our data set consisted of 1090 images. After the augmentation, the data set increased to 8720 images. Detailed information on samples per class is shown in Table 6 . Table 6 Number of images on training, validation, and test sets after applying data augmentation. Class Name Before augmentation (n) After augmentation (n) Training + validation (n) Test (n) Closed wound 141 1128 1032 96 Crust 389 3112 2792 320 Erosion 100 800 728 72 Fissure 317 2536 2336 200 Intact 143 1144 1048 96 Model Development Architecture With the learnings from our previous study, which examined several CNNs on a partially similar dataset [ 21 ], the Resnet50 [ 40 ] model showed superior performance when classifying breastfeeding-related conditions. Therefore, we apply the Resnet50 model to classify images in our dataset in two independent algorithms: one for binary classification of damage detection (i.e., informing whether damage is present) and the other to classify which damage is present in the image. Figure 5 shows the architecture of the Resnet50 model. The models were developed in Python, where we used the PyTorch library for image classification, in which the model had all layers frozen except for the last layer, which was replaced with a fully connected layer adapted to the number of classes—2 for binary damage detection and 4 for the damage classification task. Training, Validation, and Testing All models were trained for 100 epochs using the AdamW optimizer with a learning rate of 3e-4, weight decay of 0.1, and batch size of 16. We chose 100 epochs because it was a converging point where the accuracy no longer increased or decreased. For the loss functions, we applied Binary Cross-Entropy with Logits Loss for binary classification tasks, and for multiclass tasks, we used Cross-Entropy Loss, both fine-tuned with class weights to strategically adjust for class imbalances by proportionally penalizing misclassifications in less represented classes. The models were evaluated using stratified k-fold cross-validation with 10 folds. To ensure the robustness of our cross-validation process, we reset any learned parameters by initializing the models from scratch at the beginning of each fold. Instead of using the entire image data set to train the model, we did feature extraction to optimize the training process (detailed in the Feature Extraction section). We used the NVIDIA A100 SXM4 GPU architecture with 40 GB RAM to evaluate the model’s performance. After applying data augmentation to upsample the dataset, the 7936 images were allocated for training and validation, split using stratified k-fold cross-validation [ 41 ] with 10 folds, represented in Fig. 6 . In this process, 90% (7142/7936) of the data are used for training and 10% (794/7936) for validation within each fold. The stratified k-fold maintains the proportion of images in each class in both train and validation splits, ensuring each fold represents the overall data set. The remaining 784 images were left for the independent test set and were excluded entirely from these folds. We reserved the test set exclusively for final testing to assess the model’s performance on unseen data. Similarly, we performed an upsample on the test set to maintain consistency with the expanded training and validation data, ensuring the model’s evaluation of unseen examples remains robust. Since we evaluate the model in two tasks, one binary and one multiclass, the distribution of images per task is as follows: (1) for the binary damage detection, we grouped the damage classes (closed wound, crust, erosion and fissure) into one class called damage, which contains 6888 lesion images in the training and validation set and 688 in the test set. The intact class was maintained with 1048 intact images in the training and validation sets and 96 images in the test set. (2) for the four-class lesion classification, the distribution of images is the same as described in Table 4 , except that only the classes containing damages were used, removing the intact from this task. We performed feature extraction using the Resnet50 model pre-trained on the ImageNet data set. This process helped reduce the computational resources necessary for processing the data set by transforming images into numerical features without losing relevant information, helping accelerate the algorithm execution time. The models were set to evaluation mode, in which the feature maps are extracted from the final convolutional layers. These maps were then processed through adaptive pooling and flattened into 1D arrays. The extracted features were saved and used as input for the model classifiers. Tasks and Performance Metrics As the AI Algorithms section mentioned, one CNN was trained on the data set. We proposed two tasks in this study, which evaluate the CNN in the following data sets: (1) multiclass augmented and (2) binary augmented. The binary model was implemented to assess the model’s capacity to differentiate between damaged and intact images. The model underwent k-fold cross-validation, where we collected performance metrics from each fold and computed their average. We assessed the model’s performance for the multiclass data set using the following metrics: accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). In addition to the performance metrics, we generated the class activation maps using the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm [ 42 ] to understand which image features the model used to classify the damages. Declarations Author Contribution JDS, KPC, and EJW conceived and designed the study. BTSC, KC, DMK, and ACFVA assisted in the study design and supervised data collection. KPC, BTSC, and KC analyzed and interpreted the input data. JDS and EJW designed and developed the model and preprocessing pipeline. JDS, EJW, and KPC evaluated model performance. JDS, KPC, and EJW drafted the manuscript. All authors contributed to providing critical feedback and reviewed the final manuscript for important intellectual content. Acknowledgement The authors thank the Google Health Equity Research Initiative and the Merkin Graduate Fellows Program, which supported this research through their program to advance translational medicine and improve health outcomes for groups disproportionately impacted by health disparities. The authors also thank the National Council for Scientific and Technological Development (CNPq Brazil, n. 449244/2014-8), who funded the data collection portion of the study. The funder played no role in the study design, data collection, analysis, data interpretation, or this manuscript's writing. Data Availability The data that support the findings of this study are not openly available due to reasons of sensitivity and are available from the corresponding author upon reasonable request. Data are located in controlled access data storage at the Federal University of São Paulo. References Buck, M. L., Amir, L. H., Cullinane, M. & Donath, S. M. Nipple pain, damage, and vasospasm in the first 8 weeks postpartum. Breastfeeding Medicine 9, 56–62 (2014). https://doi.org/10.1089/bfm.2013.010 Koberling, A., Kopcik, K., Koper, J., Bichalska-Lach, M. & Rudzki, M. Nipple trauma in lactation - literature review. Journal of Pre-Clinical and Clinical Research. 17, 171–175 (2023). https://doi.org/10.26444/jpccr/170191 Carreiro, J. de et al. Dificuldades relacionadas Ao Aleitamento Materno: Análise de um serviço especializado em amamentação. Acta Paulista de Enfermagem 31, 430–438 (2018). https://doi.org/10.1590/1982-0194201800060 Camargo, B. T., Sañudo, A., Kusahara, D. M. & Coca, K. P. Initial nipple damages in breastfeeding women: Analysis of photographic images and Clinical Associations. Revista Brasileira de Enfermagem 77, (2024). https://doi.org/10.1590/0034-7167-2022-0773 Odom, E. C., Li, R., Scanlon, K. S., Perrine, C. G. & Grummer-Strawn, L. Reasons for earlier than desired cessation of breastfeeding. Pediatrics 131, (2013). https://doi.org/10.1542/peds.2012-1295 De Paula Leite, C. C., Mittang, B. T. & Giovanini Rossetto, E. Fatores de Risco Para Interrupção do Aleitamento Materno Exclusivo no Primeiro mês de Vida. Journal of Nursing and Health 14, (2024). https://doi.org/10.15210/jonah.v14i1.25559 Douglas, P. Re-thinking lactation-related nipple pain and damage. Women’s Health 18, (2022). https://doi.org/10.1177/17455057221087865 Anstey, E. H. et al. Lactation consultants’ perceived barriers to providing professional breastfeeding support. Journal of Human Lactation 34, 51–67 (2017). https://doi.org/10.1177/0890334417726305 Coca, K. P. et al. Measurement tools and intensity of nipple pain among women with or without damaged nipples: A quantitative systematic review. Journal of Advanced Nursing 75, 1162–1172 (2019). https://doi.org/10.1111/jan.13908 Cervellini, M. P., Gamba, M. A., Coca, K. P. & Abrão, A. C. Injuries resulted from breastfeeding: A new approach to a known problem. Revista da Escola de Enfermagem da USP 48, 346–356 (2014). https://doi.org/10.1590/S0080-6234201400002000021 Cirico, M. O., Shimoda, G. T. & Oliveira, R. N. Qualidade assistencial EM Aleitamento Materno: Implantação do indicador de trauma mamilar. Revista Gaúcha de Enfermagem 37, (2016). https://doi.org/10.1590/1983-1447.2016.04.60546 Cervellini, M. P., Coca, K. P., Gamba, M. A., Marcacine, K. O. & Abrão, A. C. Construction and validation of an instrument for classifying nipple and areola complex lesions resulting from breastfeeding. Revista Brasileira de Enfermagem 75, (2022). https://doi.org/10.1590/0034-7167-2021-0051 Nakamura, M., Asaka, Y., Ogawara, T. & Yorozu, Y. Nipple skin trauma in breastfeeding women during postpartum week one. Breastfeeding Medicine 13, 479–484 (2018). https://doi.org/10.1089/bfm.2017.021 Jesus, P. C., Oliveira, M. I. & Moraes, J. R. Capacitação de Profissionais de Saúde em Aleitamento materno e sua associação com conhecimentos, habilidades E PRÁTICAS. Ciência & Saúde Coletiva 22, 311–320 (2017). https://doi.org/10.1590/1413-81232017221.17292015 Ekem, L., Skerrett, E., Huchko, M. & Ramanujam, N. Automated Image Clarity Detection for the improvement of colposcopy imaging with multiple devices. Biomedical Signal Processing and Control 100, 106948 (2025). https://doi.org/10.1016/j.bspc.2024.106948 Bogaerts, J. M. et al. Deep learning detects premalignant lesions in the fallopian tube. npj Women’s Health 2, 11 (2024). https://doi.org/10.1038/s44294-024-00016-0 Aldhyani, T. H., Nair, R., Alzain, E., Alkahtani, H. & Koundal, D. Deep learning model for the detection of real time breast cancer images using improved dilation-based method. Diagnostics 12, 2505 (2022). https://doi.org/10.3390/diagnostics12102505 Kim, S.-Y. et al. Deep learning-based computer-aided diagnosis in screening breast ultrasound to reduce false-positive diagnoses. Scientific Reports 11, 395 (2021). https://doi.org/10.1038/s41598-020-79880-0 Lin, Z. et al. Deep learning with information fusion and model interpretation for long-term prenatal fetal heart rate data. npj Women’s Health 2, 31 (2024). https://doi.org/10.1038/s44294-024-00033-z Chiou, N., Young-Lin, N., Kelly, C. et al. Development and evaluation of deep learning models for cardiotocography interpretation. npj Womens Health 3, 21 (2025). https://doi.org/10.1038/s44294-025-00068-w De Souza, J., Viswanath, V. K., Echterhoff, J. M., Chamberlain, K. & Wang, E. J. Augmenting telepostpartum care with vision-based detection of breastfeeding-related conditions: Algorithm development and validation. JMIR AI 3, e54798 (2024). https://doi.org/10.2196/54798 De Souza, J., Chamberlain, K. & Wang, E. J. LCBuddy: Towards a smartphone-based self-assessment tool for postpartum patients with breast pain. Extended Abstracts of the CHI Conference on Human Factors in Computing Systems 1–7 (2024). https://dl.acm.org/doi/10.1145/3613905.3650944 Nakamura, M., Sugimori, H. & Ebina, Y. Development of nipple trauma evaluation system with Deep Learning. Journal of Human Lactation 41, 105–114 (2024). https://doi.org/10.1177/08903344241303867 Harris, D. & Harris, S. Digital Design and Computer Architecture 129 (Morgan Kaufmann, San Francisco, 2012). Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011). Wilson-Clay, B. & Hoover, K. The Breastfeeding Atlas . (LactNews Press, 2022). de Souza, J., Calsinski, C., Chamberlain, K., Cibrian, F. & Wang, E. J. Investigating interactive methods in remote chestfeeding support for lactation consulting professionals in Brazil. Frontiers in Digital Health 5, 1143528 (2023). https://doi.org/10.3389/fdgth.2023.1143528 Abou-Dakn, M., Fluhr, J. W., Gensch, M. & Wöckel, A. Positive effect of HPA lanolin versus expressed breastmilk on painful and damaged nipples during lactation. Skin Pharmacol. Physiol. 24, 27–35 (2011). https://doi.org/10.1159/000318228 Niazi, A. et al. A systematic review on prevention and treatment of nipple pain and fissure: Are they curable? Journal of Pharmacopuncture 21, 139–150 (2018). https://doi.org/10.3831/kpi.2018.21.017 Mitoulas, L. R. & Davanzo, R. Breast pumps and mastitis in breastfeeding women: Clarifying the relationship. Frontiers in Pediatrics 10, 856353 (2022). https://doi.org/10.3389/fped.2022.856353 Coca, K. P. et al. Efficacy of low-level laser therapy in relieving nipple pain in breastfeeding women: A triple-blind, randomized, controlled trial. Pain Management Nursing 17, 281–289 (2016). https://doi.org/10.1016/j.pmn.2016.05.003 Camargo, B.T.S., Coca, K.P., Amir, L.H. et al. The effect of a single irradiation of low-level laser on nipple pain in breastfeeding women: a randomized controlled trial. Lasers Med Sci 35, 63–69 (2020). https://doi.org/10.1007/s10103-019-02786-5 Camargo, B.T.S. Nível de dor, regeneração tecidual e carga microbiana: estudo multifásico com lesões mamilares decorrentes da amamentação [Doctoral Dissertation]. 1–174 (Escola Paulista de Enfermagem, Universidade Federal de São Paulo, 2024). https://repositorio.unifesp.br/items/ed1273eb-3a 91-4f80-9d6d-c326d49aa5c6 Stone, K., Wheeler, A. A Review of Anatomy, Physiology, and Benign Pathology of the Nipple. Ann Surg Oncol 22, 3236–3240 (2015). https://doi.org/10.1245/s10434-015-4760-4 British Medical Journal Publishing Group. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385, q902 (2024). https://doi.org/10.1136/bmj.q902 Rafay, A. & Hussain, W. Efficientskindis: An EfficientNet-based classification model for a large manually curated dataset of 31 skin diseases. Biomedical Signal Processing and Control 85, 104869 (2023). https://doi.org/10.1016/j.bspc.2023.104869 Vodrahalli, K. et al. TrueImage: A machine learning algorithm to improve the quality of telehealth photos. Biocomputing 2021, 220–231 (2021). https://doi.org/10.1142/9789811232701_0021 Finnane, A. et al. Proposed technical guidelines for the acquisition of clinical images of skin-related conditions. JAMA Dermatology 153, 453 (2017). https://doi.org/10.1001/jamadermatol.2016.6214 Perez, F., Vasconcelos, C., Avila, S. & Valle, E. Data augmentation for skin lesion analysis. In Proceedings of the Third International Skin Imaging Collaboration Workshop (ISIC 2018, Granada, Spain, September 16–20, 2018). https://doi.org/10.1007/978-3-030-01201-4_33 He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016, Las Vegas, NV, June 27–30, 2016). https://doi.org/10.1109/CVPR.2016.90 Kohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model selection. Proc. 14th Int. Joint Conf. Artif. Intell. 2, 1137–1143 (1995). https://dl.acm.org/doi/10.5555/1643031.1643047 Selvaraju, R.R., Cogswell, M., Das, A. et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Int J Comput Vis 128, 336–359 (2020). https://doi.org/10.1007/s11263-019-01228-7 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 29 Aug, 2025 Reviews received at journal 09 Aug, 2025 Reviewers agreed at journal 18 Jul, 2025 Reviews received at journal 16 May, 2025 Reviewers agreed at journal 26 Apr, 2025 Reviewers agreed at journal 26 Apr, 2025 Reviewers invited by journal 23 Apr, 2025 Editor assigned by journal 03 Apr, 2025 Submission checks completed at journal 02 Apr, 2025 First submitted to journal 31 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6348223","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":438120142,"identity":"832e12ec-435d-41b4-a467-904ed4480d19","order_by":0,"name":"Jessica de Souza","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxUlEQVRIiWNgGAWjYBACxoYECIMfRPCABIjWItlArBYGBqgWgwPEamFuTz784WNbnbzxjeTHH94w2MhuOEDIYT3P0iRntrEZbruRZiY5hyHNmLCWGTlmzLxtPIzbbuSwMfMwHE4kQkv+58+8bRL2m2fkMH/mYfhPjJYcBmneNoPEDRJABg/DASK09Dwzk5xxLiF5xhkgY45BsvFMQloM24EB9aGszrYfxHhTYSfbR1BLAwrXgIByEJAnQs0oGAWjYBSMdAAAVWND9Wb/850AAAAASUVORK5CYII=","orcid":"","institution":"University of California, San Diego","correspondingAuthor":true,"prefix":"","firstName":"Jessica","middleName":"","lastName":"de Souza","suffix":""},{"id":438120144,"identity":"b22b34e5-77d5-46b4-a566-01d465674e96","order_by":1,"name":"Kelly Pereira Coca","email":"","orcid":"","institution":"Federal University of São Paulo","correspondingAuthor":false,"prefix":"","firstName":"Kelly","middleName":"Pereira","lastName":"Coca","suffix":""},{"id":438120146,"identity":"77202b84-1371-417e-97e2-3784bc7aae93","order_by":2,"name":"Bárbara Tideman Sartorio Camargo","email":"","orcid":"","institution":"Federal University of São Paulo","correspondingAuthor":false,"prefix":"","firstName":"Bárbara","middleName":"Tideman Sartorio","lastName":"Camargo","suffix":""},{"id":438120147,"identity":"b50ef975-cd0a-449c-af89-073fad1a8a40","order_by":3,"name":"Kristina Chamberlain","email":"","orcid":"","institution":"University of California, San Diego","correspondingAuthor":false,"prefix":"","firstName":"Kristina","middleName":"","lastName":"Chamberlain","suffix":""},{"id":438120149,"identity":"c82368d5-5192-4228-ba7b-f79f915e70e1","order_by":4,"name":"Ana Cristina Freitas de Vilhena Abrão","email":"","orcid":"","institution":"Federal University of São Paulo","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"Cristina Freitas de Vilhena","lastName":"Abrão","suffix":""},{"id":438120150,"identity":"35a94935-64c8-405b-a9f0-ccfb3a7dc182","order_by":5,"name":"Denise Miyuki Kusahara","email":"","orcid":"","institution":"Federal University of São Paulo","correspondingAuthor":false,"prefix":"","firstName":"Denise","middleName":"Miyuki","lastName":"Kusahara","suffix":""},{"id":438120151,"identity":"c52baccc-b79a-405f-b693-f23190f4d26d","order_by":6,"name":"Edward Jay Wang","email":"","orcid":"","institution":"University of California, San Diego","correspondingAuthor":false,"prefix":"","firstName":"Edward","middleName":"Jay","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2025-04-01 00:38:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6348223/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6348223/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":80145816,"identity":"c0540b83-16c7-44a6-a8d2-1946231cd854","added_by":"auto","created_at":"2025-04-08 12:28:50","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":259074,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel performance on binary classification. \u003c/strong\u003e(a) Aggregated confusion matrix for the Resnet50 model. (b) Receiver operating characteristic (ROC) curve for the model performance on the test set. (c) Distribution of predicted probabilities per class is illustrated through the density plot and boxplot, respectively.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/99ac8414cf1cecee9469f4c1.png"},{"id":80146073,"identity":"6cb2e2c2-152c-40c1-b525-d17d097949d3","added_by":"auto","created_at":"2025-04-08 12:36:50","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":352405,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel performance on multi-class classification. \u003c/strong\u003e(a) Aggregated confusion matrix for the Resnet50 model. (b) Receiver operating characteristic (ROC) curve for the model performance on the test set. (c) Distribution of predicted probabilities per class is illustrated through the density plot and boxplot, respectively.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/5c7ef17ae10d7f355fea8f7c.png"},{"id":80146075,"identity":"2de96e18-4893-4317-b432-03515a5eed27","added_by":"auto","created_at":"2025-04-08 12:36:50","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2072868,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTrue positives and false negative examples. \u003c/strong\u003eImage examples from each class on the test set with its respective GRAD-CAM activation maps. The blue hues in the heatmap indicate areas with lower relevance for the model, while red hues represent higher relevance for the model to predict the class of interest.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/67d48b8d49010b34de01ab57.png"},{"id":80145818,"identity":"14d0e51c-7845-4c62-b036-58eb9049663a","added_by":"auto","created_at":"2025-04-08 12:28:50","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":450998,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDataset preprocessing pipeline.\u003c/strong\u003e Due to different image angles, a manual Region of Interest (ROI) selection was made for optimal nipple and damage area selection. Images were cropped at a 1:1 ratio and upsampled through data augmentation techniques, where we augmented the dataset from 1090 images to 8720 images.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/cca3ad3669adf1f4f1c4cb83.png"},{"id":80146072,"identity":"49ba7c4a-aa04-4495-90bf-f0616ab02628","added_by":"auto","created_at":"2025-04-08 12:36:50","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":59709,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrediction model.\u003c/strong\u003e Resnet50 model architecture for the damage classification task.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/b9e51b4994db49e43712d41e.png"},{"id":80146074,"identity":"c74cecd4-3ede-403a-8fa2-734283d5748e","added_by":"auto","created_at":"2025-04-08 12:36:50","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":73709,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStratified cross-validation.\u003c/strong\u003e Flowchart representation of the K-fold cross-validation algorithm for the ResNet50 model.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/3aae364e404e404ea428c4ed.png"},{"id":80147438,"identity":"3d5b0e2a-6993-43b1-a140-aa286b2c3b75","added_by":"auto","created_at":"2025-04-08 12:52:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4372266,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6348223/v1/6dc5d4d4-4931-4cd8-a1b9-10201ad607f9.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Automated Detection and Classification of Nipple Damage in Lactation Care","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLactation-related nipple damage is a common condition affecting breastfeeding mothers, with studies estimating that 13\u0026ndash;92% experience some degree of this complication within the first three weeks postpartum, often leading to early breastfeeding cessation [\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Causes of nipple damage include improper latching and positioning of the infant onto the breast, infections, infant anatomic anomaly (i.e., tongue-tie, high palate), and nipple anatomy [\u003cspan additionalcitationids=\"CR5 CR6\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Once a professional identifies the nipple damage, assesses its cause, and proposes a treatment, it can take several days to heal based on its initial size and depth [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Proper nipple damage assessments by health professionals are crucial for effective and timely treatments, providing patients with adequate healing time that fits the maintenance of breastfeeding [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Misdiagnosis of nipple damage and lack of consensus on treatment between health professionals can prolong treatment times, burden the patient with physical and mental pain, cause infant distress, and put the continuity of breastfeeding at risk [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCurrent methods for nipple damage classification typically involve detailed clinical examination, visual inspection with accessories for better visualization (i.e., direct light, magnifying glass), measuring tools and scales, and digital image acquisition for clinical discussion and monitoring damage progression [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Classifying nipple damage by type is a time-consuming and subjective process, often leading to variability in assessments by different health providers [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Recent advancements in nipple damage research have brought standardization of damage assessment, such as the \u003cem\u003e\u0026ldquo;Classification Instrument for Nipple-Areolar Lesions\u0026rdquo;\u003c/em\u003e (in its original naming in Brazilian Portuguese: Instrumento de Classifica\u0026ccedil;\u0026atilde;o das Les\u0026otilde;es Mamilo-Areolares) (\u003cem\u003eILMA\u003c/em\u003e) by Cervellini et al. 2022 [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] and the instrument \u003cem\u003e\u0026ldquo;Seven Signs of Nipple Trauma Associated with Breastfeeding\u0026rdquo;\u003c/em\u003e from Nakamura et al. 2018 [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. These studies emphasize the importance of proper terminology of damage under a dermatological approach, as nipple damage is frequently misclassified under the broad umbrella of 'fissures' without accounting for distinct lesion characteristics [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. This happens due to a lack of consensus in the literature on nipple injury nomenclature [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] and reduced specialized lactation training, as the number of board-certified IBCLCs is still rising [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Classifying the damage by type and damage degree, as well as identifying the cause, can contribute to a more specific treatment and bring more satisfactory results for patients who suffer from nipple pain.\u003c/p\u003e \u003cp\u003eThe widespread adoption of Deep Learning (DL) in women\u0026rsquo;s health applications has benefited clinical decision-making in several fields and enhanced patient care for better accuracy when diagnosing conditions, ranging from cervical and breast cancer exam assessments [\u003cspan additionalcitationids=\"CR16 CR17\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] to maternal and fetal well-being predictions [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Only recently, DL was adopted in lactation care to identify breastfeeding-related conditions. Firstly, Convolutional Neural Networks (CNNs) were applied to breast and nipple images to detect different lactation complications, such as abscess, mastitis, nipple blebs, dermatosis, engorgement, and nipple damage [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. De Souza et al. used a dataset of 1078 images and applied data augmentation techniques, achieving an area under the receiver operating characteristics curve (AUROC) of 0.93 for the best model. This demonstrated the potential of algorithms to assist healthcare providers in identifying complications and their applicability for at-home tele-triaging in the first few weeks postpartum [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Another example of DL in lactation care is the work from Nakamura et al., which employed a nipple damage evaluation system using their previous measurement tool, \u003cem\u003e\u0026ldquo;Seven Signs of Nipple Trauma Associated with Breastfeeding,\u0026rdquo;\u003c/em\u003e which used a dataset of 753 images and achieved an average AUROC above 0.7 to classify the classes erythema, swelling, blistering, scabbing, fissure, purpura, peeling and none [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. These previous studies highlighted the potential of using new technologies to address critical inequities in lactation care access and improve outcomes for mothers where there is a lack of professional availability and expertise. However, these studies also uncovered several challenges in classifying breastfeeding complications and nipple damage, mainly related to limited data, such as small image samples for using more robust DL algorithms, lack of racial and geographic diversity due to data collection limitations, which may not generalize well in diverse ethnic groups, and imbalanced samples in categories that are harder to classify between, that is also a problem in the real-world [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study addresses previous work gaps by developing an automated system for detecting and classifying nipple damage/lesions using DL techniques. We build on previous research by using a dataset from clinical trials in S\u0026atilde;o Paulo, Brazil, and online gathered images from a previous study. This study focuses on applying a ResNet50 Convolutional Neural Network (CNN) to perform binary and multiclass classification tasks, helping clinicians differentiate between healthy nipples and various types of nipple damage. This work also introduces a novel approach by incorporating a racially diverse dataset to ensure the model's broader applicability, especially in Latin-American contexts.\u003c/p\u003e \u003cp\u003eOur main contributions are listed as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWe present the feasibility of using DL models to detect and classify nipple damage using image classification techniques while providing a detailed methodology for application reproducibility.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe evaluate the model\u0026rsquo;s ability to generalize to unseen data by applying stratified 10-fold cross-validation, reducing bias in our metrics, and maximizing data use.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe propose end-to-end image preprocessing, data augmentation techniques, and evaluation methods to overcome challenges when handling a limited dataset.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe discuss the implications of automated damage classification systems in real-world settings and provide insights into future designs for adequately applying the system in clinical settings.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eOverview\u003c/h2\u003e \u003cp\u003eWe evaluated the ability of the Resnet50 model to classify nipple images across two tasks. The first task was the binary classification to assess whether the images had damage. The second task was a multi-class classification of the type of damage across four classes. The dataset used had 7936 images in the training and validation set and 784 images on the independent test set for the binary classification task. We used the same dataset for the second task but removed the intact images, totaling 6888 images for the training and validation set and 688 images for the independent test set. As part of evaluating both tasks, we implemented stratified 10-fold cross-validation to test the model\u0026rsquo;s generalization ability within the imbalanced dataset. There was no hyperparameter tuning between each fold, and all tasks used the same optimizer, learning rate, weight decay, and batch size. We share the model\u0026rsquo;s performance in the binary and multiclass classifications in the following subsections. We present the following performance metrics, divided into overall average and per fold: training and validation accuracy, false negatives (FN), false positives (FP), precision, recall, and F1-score. The macro-averaged metrics are reported after concatenating the inferences in the test set from all folds. In our model, we defined equal importance in detecting each class, even with an imbalanced dataset, therefore, a macro average was appropriate. Additionally, we present figures with additional metrics for the two tasks: The Area Under the Receiver Operating Characteristic Curve (AUROC), the Receiver Operating Characteristic (ROC) curve, and the Kernel Density Estimate (KDE) plot and its respective boxplot with the distribution of predicted probabilities. To enhance our understanding of the model outputs, we employ an aggregated confusion matrix that consolidates the predictions across all 10 iterations applied to the data set. We achieved this aggregation by taking the median predicted class for each instance over the multiple folds, synthesizing a singular prediction representing the agreement of the model\u0026rsquo;s behavior across the independent test set.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEvaluation of Model’s Performance on Binary Classification\u003c/h3\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the results for 10-fold cross-validation, demonstrating the model\u0026rsquo;s consistent performance across the iterations. In Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, we present several performance metrics from the model: Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea shows the aggregated confusion matrix for the Resnet50 model. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb displays the AUROC curve for each class, where the damage class has an AUROC of 0.99 (95% CI, 0.99 to 1.00). Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec shows the KDE plot and the boxplot with the distribution of predicted probabilities for both classes. We can observe minimal overlap between the classes in the distribution, where the probability below 0.50 corresponds to the class with lesion, and the probability above 0.5 corresponds to the intact class.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of 10-fold cross-validation for the augmented data set on the binary class damage detection task using the Resnet50 model.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10-fold iterations\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.981\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.923\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.951\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 3\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.917\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.947\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 4\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.984\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 5\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.977\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.922\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.978\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.917\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 8\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.923\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.963\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.942\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 9\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.984\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.917\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 10\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.987\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.928\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.947\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eBased on the aggregated confusion matrix, out of the 784 images used in the testing set, the model could correctly classify 762 images, giving a recall of 95.90%, a precision of 91.72%, and an F1-score of 93.67%. The total images correctly classified by category are as follows: damage is the best class (671/688; accuracy\u0026thinsp;=\u0026thinsp;98%, F1-score\u0026thinsp;=\u0026thinsp;0.984), followed by intact (91/96; accuracy\u0026thinsp;=\u0026thinsp;95%, F1-score\u0026thinsp;=\u0026thinsp;0.890). Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes the model\u0026rsquo;s performance per class on the augmented test set.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of the detection results per class: accuracy, precision, recall, F1-score, and support (ie, number of samples per class) using the Resnet50 architecture.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass Name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFP (N)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF\u003csub\u003e1\u003c/sub\u003e-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupport\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDamage\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.992\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.975\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.984\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e688\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntact\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eEvaluation of Model’s Performance on Multiclass Classification\u003c/h3\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the results for 10-fold cross-validation on the multiclass classification task, highlighting the model\u0026rsquo;s consistent performance across the iterations. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea shows the aggregated confusion matrix from the Resnet50 model. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb shows the AUROC curve for each class, with the following values: closed wound (AUROC\u0026thinsp;=\u0026thinsp;0.98, 95% CI, 0.98 to 0.98), crust (AUROC\u0026thinsp;=\u0026thinsp;0.90, 95% CI, 0.90 to 0.91), erosion (AUROC\u0026thinsp;=\u0026thinsp;0.99, 95% CI, 0.99 to 0.99), and fissure (AUROC\u0026thinsp;=\u0026thinsp;0.89, 95% CI, 0.89 to 0.90). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec shows the KDE plot and the boxplot illustrating the distribution of predicted probabilities for each class. To estimate which class was predicted in the test set, we used a one-hot encoder [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] from the sklearn library [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], which evaluated the output probabilities for each class and binarized the highest probability as one and the lowest probabilities as zero.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of 10-fold cross-validation for the augmented data set on the multi-class damage classification task using the Resnet50 model.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10-fold iterations\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.863\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.844\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.916\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.835\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 3\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.912\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 4\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.842\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 5\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.914\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.820\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.826\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.848\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.854\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.850\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.828\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.830\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 8\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.842\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 9\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIteration 10\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.919\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.828\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.848\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eBased on the aggregated confusion matrix, out of the 688 images used in the testing set, the model could correctly classify 542 images, giving a macro-average precision of 76.83%, recall of 82.07%, and F1-score of 78.96%. The total images correctly classified by category are as follows: closed wound (88/96; accuracy\u0026thinsp;=\u0026thinsp;92%), crust (250/320; accuracy\u0026thinsp;=\u0026thinsp;78%), erosion (65/72; accuracy\u0026thinsp;=\u0026thinsp;90%), and fissure (139/200; accuracy\u0026thinsp;=\u0026thinsp;69%). Among the remaining images that were misclassified with a percentage of error higher than 10%, the class crust had 12% (37/320) of images misclassified as fissure, the class fissure had 21% (42/200) of images misclassified as crusts, and the class erosion had 10% (7/72) misclassified as crusts. The best-performing class was erosion (F1-score\u0026thinsp;=\u0026thinsp;0.829), followed by closed wound (F1-score\u0026thinsp;=\u0026thinsp;0.809), crust (F1-score\u0026thinsp;=\u0026thinsp;0.803), and fissure (F1-score\u0026thinsp;=\u0026thinsp;0.718). Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e summarizes the model\u0026rsquo;s performance per class on the augmented test set.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of the detection results per class: accuracy, precision, recall, F1-score, and support (ie, number of samples per class) using the Resnet50 architecture.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass Name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFN (N)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFP (N)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF\u003csub\u003e1\u003c/sub\u003e-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSupport\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eClosed wound\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.914\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCrust\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.830\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.779\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e320\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eErosion\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.765\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.907\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFissure\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.757\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.683\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.718\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e200\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows image samples for true positives and false negatives the original image accompanied by the image regions responsible for the model\u0026rsquo;s decision-making process through the GRAD-CAM activation maps. By overlaying the activation map on top of the original image, we can observe how the model highlights the same area corresponding to the damage and region of interest (highlighted in cyan in the original image). Each of the four classes with damages have different patterns in the activation map, where the warmest colors show areas with higher probability in the class of interest.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study evaluated the potential of DL in identifying nipple damage through image classification using an adaptation of the ILMA tool [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], which used two independent models to detect a damage through the binary task and classify the type of damage through the multiclass task. Through comprehensive testing on the independent test set, we demonstrate the robust performance and the ability of our model to distinguish between healthy and unhealthy images, as well as identify four types of nipple damage, such as damages without interruption of the skin barrier (closed wound), crusts, erosions, and fissures with high precision, recall, and F1-score. It is essential to note the model's excellent ability in the binary classification task. By identifying whether there are signs of damage in a nipple image evaluation, this model alone can help triage and indicate patients who potentially are facing some disruption and discomfort in their breastfeeding journey. Meanwhile, classifying the damage by type can help health professionals with more accurate diagnoses, allowing targeted treatments for improved healing outcomes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWhen analyzing the misclassified images in the independent test set, several factors were identified that may have contributed to the model's confusion. These include: (a) Similar visual features between classes, such as overlapping characteristics of different damage types. (b) Variations in nipple-areolar color and anatomy can affect the model's ability to distinguish between classes. (c) Environmental lighting during image acquisition, particularly in images taken with low-light conditions, where shadows on darker areolas resembled crusting in fissure images. (d) Images showing multiple damages simultaneously, such as a crust image with micro-fissures, complicate the classification process, and (e) conditions that precede or evolve into one another, such as erosion healing into a crust over time. These challenges were also mentioned in previous work and are not just technical limitations, as they also happen in real-world settings and can impact professional decision-making [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo provide further insight into the binary classification task, images with closed wound damages were occasionally misclassified as intact and vice versa, particularly in lighter skin tones and pink areolas. Similarly, confusion between crust and erosion damages arose due to the color variations in scabbing that varied (from red, light yellow to brown), overlapping with the red, fluid-filled appearance of some erosion images. Additionally, some images were taken over several days with intervals of 12 hours, capturing the transition of an erosion damage healing into a crust, which may have further complicated the model's ability to differentiate between the two. Another challenge was the presence of nipple anatomy variation, e.g., inverted and protruding nipples with a granular texture, resembling a \"raspberry nipple\" [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], which could be mistaken for vesicles or blistering despite being healthy. While the model may have learned certain features from the same user across different images, variations in lighting and angles likely influenced its performance. Although data augmentation techniques helped mitigate overfitting and increased image variety, some issues persisted. However, these challenges do not significantly impact the model's overall performance, particularly in cases where damages naturally evolve, such as an erosion transitioning to a crust.\u003c/p\u003e \u003cp\u003eThe ability to classify nipple damage using DL with the ILMA classification tool offers significant potential to standardize and promote the use of proper terminologies, fostering greater adoption among healthcare professionals. To maximize the benefits of this technology, health specialists should receive targeted education on nipple damage classification to ensure accurate diagnosis and streamlined patient treatment [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. This algorithm has several practical applications: (1) Professional Training: It can be an educational tool to train healthcare providers on the diverse types of nipple damage, improving their diagnostic skills and confidence. (2) Support in Underserved Areas: In regions with limited access to specialized lactation care, the algorithm can assist general health providers in accurately classifying damages and recommending appropriate treatments. (3) Remote Classification: The tool can enable remote damage classification in healthcare facilities with limited professional availability, ensuring timely and accurate assessments. (4) Patient Triage and Hybrid Care: For lactation providers offering hybrid care models, the algorithm can help prioritize patients based on damage severity and streamline workflows [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. (5) Postpartum Monitoring: When integrated with ecological momentary assessment (EMA), the tool can facilitate regular checkups during the first few weeks postpartum, aiding in early damage recognition and prevention [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSome limitations of this work include the limited dataset size in specific classes, such as within the closed wounds group, limiting the model\u0026rsquo;s ability to separate the categories that have no skin barrier interruption, such as erythema, edema, vesicle, and ecchymosis and lead to the grouping of images into one condition. Another limitation was the lack of clinical standardization for conditions by severity to a more specialized classification. While previous work proposed a severity-based nipple trauma classification (i.e., none, minor, moderate, and severe) [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], the clinical adoption of such tiers remains limited due to the absence of standardized, validated criteria, such as using the Nipple Trauma Score (NTS) tool proposed by Abou-Dakn et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Measuring lesion severity is a subjective task, leading to the lack of professional agreement on the level of damage severity in real-world scenarios [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Given these challenges, our model focused on lesion detection and classification rather than severity grading, which is aligned with current clinical practices [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFuture works include (1) assessing the dataset images using the NTS tool across and performing a DL evaluation to assess the capacity for automated NTS detection, (2) validating a practical tool for real-time classification across multidisciplinary providers while evaluating model accuracy, professional trust, comfort, and privacy concerns in using technology to assist in their work. And (3), expand the dataset to incorporate higher racial diversity while increasing the distribution of different conditions and including additional variables such as pain levels, persistence of the damage, and treatments applied to follow damage progression.\u003c/p\u003e \u003cp\u003eIn conclusion, our study introduces an innovative DL model capable of automatically detecting and classifying nipple damage using the ILMA classification tool. The accurate assessment of breast conditions is crucial, as pain and discomfort during breastfeeding represent significant barriers for parents striving to continue breastfeeding their infants [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. We hope this work will open meaningful discussions on how technology can address critical inequities in access to lactation care and support professional education on the topic, helping to standardize the nomenclature of nipple damage while improving treatment outcomes. By supporting healthcare providers in regions with limited professional availability and expertise, this tool has the potential to improve patient outcomes, reduce the distress caused by nipple damage, and ultimately enhance the breastfeeding experience for mothers worldwide.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStudy Design\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eEthical Considerations\u003c/h2\u003e \u003cp\u003eThis multi-site study was approved by the Institutional Review Boards of the University of California, San Diego (No. 801,904) and the Federal University of S\u0026atilde;o Paulo (No. 1710/09, No. 1.267.605, and No. 4.726.329). No personally identifiable data from the participants were incorporated into this research.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eStudy Type and Data Sources\u003c/h3\u003e\n\u003cp\u003eThis study comprises two separate datasets on lactation-related nipple damage. The first dataset had the photographic images obtained through clinical settings, where participants were in the first two days postpartum, between 2011 and 2024, in S\u0026atilde;o Paulo, Brazil. The detailed protocol for participant selection and data collection is described in previous publications [\u003cspan additionalcitationids=\"CR32\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. The second dataset was gathered in a previous study and used mixed-media data collection (i.e., physical and online clinical resources) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The second dataset was included in this study to complement the clinical dataset and help with balancing classes for better algorithm generalization. We performed a retrospective analysis of the prospectively collected data before including them in the algorithm for detection and classification.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDataset Characteristics\u003c/h2\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003eParticipants\u003c/h2\u003e \u003cp\u003eA total of 1219 nipple images were collected to compose the dataset, described in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The images were categorized into primary damage type, nipple color, and damage location. Following applying the inclusion criteria, 1,090 images were retained, while 129 images with visual issues that could compromise the study were excluded. The images were categorized into five classes, following the ILMA [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] classification instrument: Closed wounds, crust, erosion, fissure, and intact. The \u003cem\u003eclosed wound\u003c/em\u003e class presents Nipple and Areola Damages (NADs) without skin barrier disruption, including erythema, ecchymosis, vesicle, and edema. This class was concatenated due to visual similarities in the conditions and the limited number of images needed to compose separate classes. The \u003cem\u003ecrust\u003c/em\u003e class shows scabbing of the skin and presents itself after the damage is healing and can have yellow, brown, or dark red tones. The \u003cem\u003eerosion\u003c/em\u003e class presents damages with well-defined borders and can contain clear or blood fluid, exposing the skin's dermis or epidermis. The \u003cem\u003efissure\u003c/em\u003e class presents damages with a narrow and linear shape, which may present varied damage depths, with or without the presence of fluids. Finally, the \u003cem\u003eintact\u003c/em\u003e class includes images without classification from the ILMA instrument, where images presenting healthy tissue and no damage were classified into this category. The classification of the nipple-areolar color presented three primary colors, where we categorized images in black, brown, and pink pigmentation [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e], with a variation between lighter and darker pigmentation that was not considered since the damages often presented themselves with redness at the nipple surface. No Fitzpatrick Skin Tone classification was performed in the images since no skin surface was evaluated in the study, only nipple-areolar tissue.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eInclusion and Exclusion Criteria\u003c/h2\u003e \u003cp\u003eTo be included in the dataset, the images had to meet the following criteria: (1) RGB format (PNG or JPEG); (2) visually present at least 1 of the 5 conditions; (3) have the nipple visible; (4) have no irrelevant objects covering the nipple; (5) have sufficient resolution for clinical evaluation, showing no signs of low quality, such as blurriness (due to motion or focus issues), pixelation, overexposure or underexposure. Images were excluded from the data set if (1) the breast or nipple was from non-lactating female patients; (2) there was no match between the image and the label; (3) the breast or nipple was not visible in the image; and (4) the image had no label describing it.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOriginal dataset characteristics.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDetails\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDataset Size (MB)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e344.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eImage Size (KB)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMin: 6.46, Average: 326.12, Max: 1570.42\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eImage Dimensions (pixels)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMin: 77 \u0026times; 77, Average: 740 \u0026times; 740, Max: 1773 \u0026times; 1773\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eData Collection Period\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2011 to 2024\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eData Source\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrior Clinical Trials in S\u0026atilde;o Paulo: 847 [\u003cspan additionalcitationids=\"CR32\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMixed-media source from previous study: 243 [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTotal Images Collected\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1219\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTotal Images Excluded\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e129 (10.58%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFinal Dataset Size\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1090 (89.42%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNumber of Classes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNumber of Unique Subjects\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e403\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003e\u003cb\u003eImages per Damage Category\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClosed wounds: 141\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCrust: 389\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eErosion: 100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFissure: 317\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIntact: 143\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003ePredominant Nipple Color\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBlack: 57\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBrown (light to darker tones): 901\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePink (light to darker tones): 132\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eNipple Damage Location\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eApex: 864\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLateral: 83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNon-present (intact): 143\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eData Labeling Protocol\u003c/h2\u003e \u003cp\u003eTwo board-certified nurse practitioners with a PhD in Nursing, specialists in maternal and lactation care, and International Board-Certified Lactation Consultants (IBCLC) with over 10 years of experience, collaboratively classified the images on the dataset to ensure that their classifications had no discrepancies. This consensus-based classification process involved the professionals reviewing each image together and discussing the most prominent damage until they reached an agreement. The damage classification was based on the ILMA instrument [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. The entire classification process took approximately six hours.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eImage Preprocessing\u003c/h2\u003e \u003cp\u003eBefore using the images as inputs for the DL models, the images were manually cropped in a 1:1 ratio to prevent image flattening or warping during resizing and loss of essential features. Most images have the nipple tissue concentrated in the center of the image, thereby focusing the model\u0026rsquo;s evaluation on the most relevant areas. We followed the guidelines proposed by similar studies on DL for skin disease and damage classification [\u003cspan additionalcitationids=\"CR36 CR37\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], which aim to objectively show the area of interest for optimized detection and reduce the risks of poorly triaged images. The image preprocessing pipeline and samples of the augmentations used are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eBefore entering the DL pipeline, we applied standard transformations in the data, starting with image resizing. In this paper, we trained, validated, and tested our data set using the Resnet50 Convolutional Neural Network (CNN). This CNN requires the input images to be resized to 224\u0026times;224 pixels. Therefore, we proceeded with the image resizing according to each model\u0026rsquo;s requirements. The last transformation step incorporates the normalization of the images, which standardizes pixel intensity across the data set. For better model generalization in this dataset, we calculated the mean and SD of all images in the data set to use in the normalization process instead of using the ImageNet data set pre-trained parameters, following previous study involving breast disease classification [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo increase the variety of the dataset while improving the model\u0026rsquo;s robustness and reducing the risk of overfitting, we implemented data augmentation techniques to upsample the images on the dataset (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). We used the following seven data augmentations previously used in data sets involving skin damage [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]: brightness, random affine, contrast, horizontal and vertical flip, center zoom, and rotation. Before data augmentation, our data set consisted of 1090 images. After the augmentation, the data set increased to 8720 images. Detailed information on samples per class is shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eNumber of images on training, validation, and test sets after applying data augmentation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass Name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBefore augmentation (n)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAfter augmentation (n)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTraining\u0026thinsp;+\u0026thinsp;validation (n)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTest (n)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eClosed wound\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1032\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCrust\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e389\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3112\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e320\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eErosion\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e728\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFissure\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e317\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2536\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2336\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e200\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntact\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e143\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1144\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1048\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eModel Development\u003c/h2\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003eArchitecture\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWith the learnings from our previous study, which examined several CNNs on a partially similar dataset [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], the Resnet50 [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] model showed superior performance when classifying breastfeeding-related conditions. Therefore, we apply the Resnet50 model to classify images in our dataset in two independent algorithms: one for binary classification of damage detection (i.e., informing whether damage is present) and the other to classify which damage is present in the image. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the architecture of the Resnet50 model. The models were developed in Python, where we used the PyTorch library for image classification, in which the model had all layers frozen except for the last layer, which was replaced with a fully connected layer adapted to the number of classes\u0026mdash;2 for binary damage detection and 4 for the damage classification task.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eTraining, Validation, and Testing\u003c/h2\u003e \u003cp\u003eAll models were trained for 100 epochs using the AdamW optimizer with a learning rate of 3e-4, weight decay of 0.1, and batch size of 16. We chose 100 epochs because it was a converging point where the accuracy no longer increased or decreased. For the loss functions, we applied Binary Cross-Entropy with Logits Loss for binary classification tasks, and for multiclass tasks, we used Cross-Entropy Loss, both fine-tuned with class weights to strategically adjust for class imbalances by proportionally penalizing misclassifications in less represented classes. The models were evaluated using stratified k-fold cross-validation with 10 folds. To ensure the robustness of our cross-validation process, we reset any learned parameters by initializing the models from scratch at the beginning of each fold. Instead of using the entire image data set to train the model, we did feature extraction to optimize the training process (detailed in the Feature Extraction section). We used the NVIDIA A100 SXM4 GPU architecture with 40 GB RAM to evaluate the model\u0026rsquo;s performance.\u003c/p\u003e \u003cp\u003eAfter applying data augmentation to upsample the dataset, the 7936 images were allocated for training and validation, split using stratified k-fold cross-validation [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e] with 10 folds, represented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. In this process, 90% (7142/7936) of the data are used for training and 10% (794/7936) for validation within each fold. The stratified k-fold maintains the proportion of images in each class in both train and validation splits, ensuring each fold represents the overall data set. The remaining 784 images were left for the independent test set and were excluded entirely from these folds. We reserved the test set exclusively for final testing to assess the model\u0026rsquo;s performance on unseen data. Similarly, we performed an upsample on the test set to maintain consistency with the expanded training and validation data, ensuring the model\u0026rsquo;s evaluation of unseen examples remains robust. Since we evaluate the model in two tasks, one binary and one multiclass, the distribution of images per task is as follows: (1) for the binary damage detection, we grouped the damage classes (closed wound, crust, erosion and fissure) into one class called damage, which contains 6888 lesion images in the training and validation set and 688 in the test set. The intact class was maintained with 1048 intact images in the training and validation sets and 96 images in the test set. (2) for the four-class lesion classification, the distribution of images is the same as described in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, except that only the classes containing damages were used, removing the intact from this task.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe performed feature extraction using the Resnet50 model pre-trained on the ImageNet data set. This process helped reduce the computational resources necessary for processing the data set by transforming images into numerical features without losing relevant information, helping accelerate the algorithm execution time. The models were set to evaluation mode, in which the feature maps are extracted from the final convolutional layers. These maps were then processed through adaptive pooling and flattened into 1D arrays. The extracted features were saved and used as input for the model classifiers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eTasks and Performance Metrics\u003c/h2\u003e \u003cp\u003eAs the AI Algorithms section mentioned, one CNN was trained on the data set. We proposed two tasks in this study, which evaluate the CNN in the following data sets: (1) multiclass augmented and (2) binary augmented. The binary model was implemented to assess the model\u0026rsquo;s capacity to differentiate between damaged and intact images. The model underwent k-fold cross-validation, where we collected performance metrics from each fold and computed their average. We assessed the model\u0026rsquo;s performance for the multiclass data set using the following metrics: accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). In addition to the performance metrics, we generated the class activation maps using the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] to understand which image features the model used to classify the damages.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJDS, KPC, and EJW conceived and designed the study. BTSC, KC, DMK, and ACFVA assisted in the study design and supervised data collection. KPC, BTSC, and KC analyzed and interpreted the input data. JDS and EJW designed and developed the model and preprocessing pipeline. JDS, EJW, and KPC evaluated model performance. JDS, KPC, and EJW drafted the manuscript. All authors contributed to providing critical feedback and reviewed the final manuscript for important intellectual content.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors thank the Google Health Equity Research Initiative and the Merkin Graduate Fellows Program, which supported this research through their program to advance translational medicine and improve health outcomes for groups disproportionately impacted by health disparities. The authors also thank the National Council for Scientific and Technological Development (CNPq Brazil, n. 449244/2014-8), who funded the data collection portion of the study. The funder played no role in the study design, data collection, analysis, data interpretation, or this manuscript's writing.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data that support the findings of this study are not openly available due to reasons of sensitivity and are available from the corresponding author upon reasonable request. Data are located in controlled access data storage at the Federal University of S\u0026atilde;o Paulo.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBuck, M. L., Amir, L. H., Cullinane, M. \u0026amp; Donath, S. M. Nipple pain, damage, and vasospasm in the first 8 weeks postpartum. Breastfeeding Medicine 9, 56\u0026ndash;62 (2014). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1089/bfm.2013.010\u003c/span\u003e\u003cspan address=\"10.1089/bfm.2013.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoberling, A., Kopcik, K., Koper, J., Bichalska-Lach, M. \u0026amp; Rudzki, M. Nipple trauma in lactation - literature review. Journal of Pre-Clinical and Clinical Research. 17, 171\u0026ndash;175 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.26444/jpccr/170191\u003c/span\u003e\u003cspan address=\"10.26444/jpccr/170191\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarreiro, J. de \u003cem\u003eet al.\u003c/em\u003e Dificuldades relacionadas Ao Aleitamento Materno: An\u0026aacute;lise de um servi\u0026ccedil;o especializado em amamenta\u0026ccedil;\u0026atilde;o. Acta Paulista de Enfermagem 31, 430\u0026ndash;438 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/1982-0194201800060\u003c/span\u003e\u003cspan address=\"10.1590/1982-0194201800060\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCamargo, B. T., Sa\u0026ntilde;udo, A., Kusahara, D. M. \u0026amp; Coca, K. P. Initial nipple damages in breastfeeding women: Analysis of photographic images and Clinical Associations. Revista Brasileira de Enfermagem 77, (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/0034-7167-2022-0773\u003c/span\u003e\u003cspan address=\"10.1590/0034-7167-2022-0773\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOdom, E. C., Li, R., Scanlon, K. S., Perrine, C. G. \u0026amp; Grummer-Strawn, L. Reasons for earlier than desired cessation of breastfeeding. Pediatrics 131, (2013). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1542/peds.2012-1295\u003c/span\u003e\u003cspan address=\"10.1542/peds.2012-1295\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Paula Leite, C. C., Mittang, B. T. \u0026amp; Giovanini Rossetto, E. Fatores de Risco Para Interrup\u0026ccedil;\u0026atilde;o do Aleitamento Materno Exclusivo no Primeiro m\u0026ecirc;s de Vida. Journal of Nursing and Health 14, (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.15210/jonah.v14i1.25559\u003c/span\u003e\u003cspan address=\"10.15210/jonah.v14i1.25559\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDouglas, P. Re-thinking lactation-related nipple pain and damage. Women\u0026rsquo;s Health 18, (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/17455057221087865\u003c/span\u003e\u003cspan address=\"10.1177/17455057221087865\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnstey, E. H. \u003cem\u003eet al.\u003c/em\u003e Lactation consultants\u0026rsquo; perceived barriers to providing professional breastfeeding support. Journal of Human Lactation 34, 51\u0026ndash;67 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0890334417726305\u003c/span\u003e\u003cspan address=\"10.1177/0890334417726305\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoca, K. P. \u003cem\u003eet al.\u003c/em\u003e Measurement tools and intensity of nipple pain among women with or without damaged nipples: A quantitative systematic review. Journal of Advanced Nursing 75, 1162\u0026ndash;1172 (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/jan.13908\u003c/span\u003e\u003cspan address=\"10.1111/jan.13908\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCervellini, M. P., Gamba, M. A., Coca, K. P. \u0026amp; Abr\u0026atilde;o, A. C. Injuries resulted from breastfeeding: A new approach to a known problem. Revista da Escola de Enfermagem da USP 48, 346\u0026ndash;356 (2014). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/S0080-6234201400002000021\u003c/span\u003e\u003cspan address=\"10.1590/S0080-6234201400002000021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCirico, M. O., Shimoda, G. T. \u0026amp; Oliveira, R. N. Qualidade assistencial EM Aleitamento Materno: Implanta\u0026ccedil;\u0026atilde;o do indicador de trauma mamilar. Revista Ga\u0026uacute;cha de Enfermagem 37, (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/1983-1447.2016.04.60546\u003c/span\u003e\u003cspan address=\"10.1590/1983-1447.2016.04.60546\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCervellini, M. P., Coca, K. P., Gamba, M. A., Marcacine, K. O. \u0026amp; Abr\u0026atilde;o, A. C. Construction and validation of an instrument for classifying nipple and areola complex lesions resulting from breastfeeding. Revista Brasileira de Enfermagem 75, (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/0034-7167-2021-0051\u003c/span\u003e\u003cspan address=\"10.1590/0034-7167-2021-0051\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNakamura, M., Asaka, Y., Ogawara, T. \u0026amp; Yorozu, Y. Nipple skin trauma in breastfeeding women during postpartum week one. Breastfeeding Medicine 13, 479\u0026ndash;484 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1089/bfm.2017.021\u003c/span\u003e\u003cspan address=\"10.1089/bfm.2017.021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJesus, P. C., Oliveira, M. I. \u0026amp; Moraes, J. R. Capacita\u0026ccedil;\u0026atilde;o de Profissionais de Sa\u0026uacute;de em Aleitamento materno e sua associa\u0026ccedil;\u0026atilde;o com conhecimentos, habilidades E PR\u0026Aacute;TICAS. Ci\u0026ecirc;ncia \u0026amp; Sa\u0026uacute;de Coletiva 22, 311\u0026ndash;320 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1590/1413-81232017221.17292015\u003c/span\u003e\u003cspan address=\"10.1590/1413-81232017221.17292015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkem, L., Skerrett, E., Huchko, M. \u0026amp; Ramanujam, N. Automated Image Clarity Detection for the improvement of colposcopy imaging with multiple devices. Biomedical Signal Processing and Control 100, 106948 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.bspc.2024.106948\u003c/span\u003e\u003cspan address=\"10.1016/j.bspc.2024.106948\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBogaerts, J. M. \u003cem\u003eet al.\u003c/em\u003e Deep learning detects premalignant lesions in the fallopian tube. npj Women\u0026rsquo;s Health 2, 11 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s44294-024-00016-0\u003c/span\u003e\u003cspan address=\"10.1038/s44294-024-00016-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAldhyani, T. H., Nair, R., Alzain, E., Alkahtani, H. \u0026amp; Koundal, D. Deep learning model for the detection of real time breast cancer images using improved dilation-based method. Diagnostics 12, 2505 (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/diagnostics12102505\u003c/span\u003e\u003cspan address=\"10.3390/diagnostics12102505\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, S.-Y. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based computer-aided diagnosis in screening breast ultrasound to reduce false-positive diagnoses. Scientific Reports 11, 395 (2021). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-020-79880-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-020-79880-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin, Z. \u003cem\u003eet al.\u003c/em\u003e Deep learning with information fusion and model interpretation for long-term prenatal fetal heart rate data. npj Women\u0026rsquo;s Health 2, 31 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s44294-024-00033-z\u003c/span\u003e\u003cspan address=\"10.1038/s44294-024-00033-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiou, N., Young-Lin, N., Kelly, C. et al. Development and evaluation of deep learning models for cardiotocography interpretation. npj Womens Health 3, 21 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s44294-025-00068-w\u003c/span\u003e\u003cspan address=\"10.1038/s44294-025-00068-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Souza, J., Viswanath, V. K., Echterhoff, J. M., Chamberlain, K. \u0026amp; Wang, E. J. Augmenting telepostpartum care with vision-based detection of breastfeeding-related conditions: Algorithm development and validation. JMIR AI 3, e54798 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2196/54798\u003c/span\u003e\u003cspan address=\"10.2196/54798\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Souza, J., Chamberlain, K. \u0026amp; Wang, E. J. LCBuddy: Towards a smartphone-based self-assessment tool for postpartum patients with breast pain. \u003cem\u003eExtended Abstracts of the CHI Conference on Human Factors in Computing Systems\u003c/em\u003e 1\u0026ndash;7 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dl.acm.org/doi/10.1145/3613905.3650944\u003c/span\u003e\u003cspan address=\"https://dl.acm.doi/10.1145/3613905.3650944\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNakamura, M., Sugimori, H. \u0026amp; Ebina, Y. Development of nipple trauma evaluation system with Deep Learning. Journal of Human Lactation 41, 105\u0026ndash;114 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/08903344241303867\u003c/span\u003e\u003cspan address=\"10.1177/08903344241303867\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarris, D. \u0026amp; Harris, S. \u003cem\u003eDigital Design and Computer Architecture\u003c/em\u003e 129 (Morgan Kaufmann, San Francisco, 2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa, F. et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, 2825\u0026ndash;2830 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilson-Clay, B. \u0026amp; Hoover, K. \u003cem\u003eThe Breastfeeding Atlas\u003c/em\u003e. (LactNews Press, 2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Souza, J., Calsinski, C., Chamberlain, K., Cibrian, F. \u0026amp; Wang, E. J. Investigating interactive methods in remote chestfeeding support for lactation consulting professionals in Brazil. Frontiers in Digital Health 5, 1143528 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fdgth.2023.1143528\u003c/span\u003e\u003cspan address=\"10.3389/fdgth.2023.1143528\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbou-Dakn, M., Fluhr, J. W., Gensch, M. \u0026amp; W\u0026ouml;ckel, A. Positive effect of HPA lanolin versus expressed breastmilk on painful and damaged nipples during lactation. Skin Pharmacol. Physiol. 24, 27\u0026ndash;35 (2011). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1159/000318228\u003c/span\u003e\u003cspan address=\"10.1159/000318228\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNiazi, A. \u003cem\u003eet al.\u003c/em\u003e A systematic review on prevention and treatment of nipple pain and fissure: Are they curable? Journal of Pharmacopuncture 21, 139\u0026ndash;150 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3831/kpi.2018.21.017\u003c/span\u003e\u003cspan address=\"10.3831/kpi.2018.21.017\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMitoulas, L. R. \u0026amp; Davanzo, R. Breast pumps and mastitis in breastfeeding women: Clarifying the relationship. Frontiers in Pediatrics 10, 856353 (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fped.2022.856353\u003c/span\u003e\u003cspan address=\"10.3389/fped.2022.856353\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoca, K. P. \u003cem\u003eet al.\u003c/em\u003e Efficacy of low-level laser therapy in relieving nipple pain in breastfeeding women: A triple-blind, randomized, controlled trial. Pain Management Nursing 17, 281\u0026ndash;289 (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.pmn.2016.05.003\u003c/span\u003e\u003cspan address=\"10.1016/j.pmn.2016.05.003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCamargo, B.T.S., Coca, K.P., Amir, L.H. \u003cem\u003eet al.\u003c/em\u003e The effect of a single irradiation of low-level laser on nipple pain in breastfeeding women: a randomized controlled trial. Lasers Med Sci 35, 63\u0026ndash;69 (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10103-019-02786-5\u003c/span\u003e\u003cspan address=\"10.1007/s10103-019-02786-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCamargo, B.T.S. N\u0026iacute;vel de dor, regenera\u0026ccedil;\u0026atilde;o tecidual e carga microbiana: estudo multif\u0026aacute;sico com les\u0026otilde;es mamilares decorrentes da amamenta\u0026ccedil;\u0026atilde;o [Doctoral Dissertation]. 1\u0026ndash;174 (Escola Paulista de Enfermagem, Universidade Federal de S\u0026atilde;o Paulo, 2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://repositorio.unifesp.br/items/ed1273eb-3a\u003c/span\u003e\u003cspan address=\"https://repositorio.unifesp.br/items/ed1273eb-3a\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e91-4f80-9d6d-c326d49aa5c6\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStone, K., Wheeler, A. A Review of Anatomy, Physiology, and Benign Pathology of the Nipple. Ann Surg Oncol 22, 3236\u0026ndash;3240 (2015). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1245/s10434-015-4760-4\u003c/span\u003e\u003cspan address=\"10.1245/s10434-015-4760-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBritish Medical Journal Publishing Group. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385, q902 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmj.q902\u003c/span\u003e\u003cspan address=\"10.1136/bmj.q902\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRafay, A. \u0026amp; Hussain, W. Efficientskindis: An EfficientNet-based classification model for a large manually curated dataset of 31 skin diseases. Biomedical Signal Processing and Control 85, 104869 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.bspc.2023.104869\u003c/span\u003e\u003cspan address=\"10.1016/j.bspc.2023.104869\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVodrahalli, K. \u003cem\u003eet al.\u003c/em\u003e TrueImage: A machine learning algorithm to improve the quality of telehealth photos. Biocomputing 2021, 220\u0026ndash;231 (2021). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1142/9789811232701_0021\u003c/span\u003e\u003cspan address=\"10.1142/9789811232701_0021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFinnane, A. \u003cem\u003eet al.\u003c/em\u003e Proposed technical guidelines for the acquisition of clinical images of skin-related conditions. JAMA Dermatology 153, 453 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1001/jamadermatol.2016.6214\u003c/span\u003e\u003cspan address=\"10.1001/jamadermatol.2016.6214\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerez, F., Vasconcelos, C., Avila, S. \u0026amp; Valle, E. Data augmentation for skin lesion analysis. In \u003cem\u003eProceedings of the Third International Skin Imaging Collaboration Workshop\u003c/em\u003e (ISIC 2018, Granada, Spain, September 16\u0026ndash;20, 2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-030-01201-4_33\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-01201-4_33\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. Deep residual learning for image recognition. In \u003cem\u003eProceedings of the IEEE Conference on Computer Vision and Pattern Recognition\u003c/em\u003e (CVPR 2016, Las Vegas, NV, June 27\u0026ndash;30, 2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/CVPR.2016.90\u003c/span\u003e\u003cspan address=\"10.1109/CVPR.2016.90\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model selection. \u003cem\u003eProc. 14th Int. Joint Conf. Artif. Intell.\u003c/em\u003e 2, 1137\u0026ndash;1143 (1995). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dl.acm.org/doi/10.5555/1643031.1643047\u003c/span\u003e\u003cspan address=\"https://dl.acm.doi/10.5555/1643031.1643047\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelvaraju, R.R., Cogswell, M., Das, A. \u003cem\u003eet al.\u003c/em\u003e Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Int J Comput Vis 128, 336\u0026ndash;359 (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11263-019-01228-7\u003c/span\u003e\u003cspan address=\"10.1007/s11263-019-01228-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-womens-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [npj Women's Health](https://www.nature.com/npjwomenshealth/)","snPcode":"44294","submissionUrl":"https://submission.springernature.com/new-submission/44294/3","title":"npj Women's Health","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6348223/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6348223/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLactation-related nipple damage is a prevalent issue among breastfeeding mothers, often leading to early breastfeeding cessation due to pain and misdiagnosis. Accurate and timely classification of nipple damage is critical for effective treatment, yet current methods rely on subjective clinical assessments, resulting in variability and inefficiency. This study addresses these challenges by developing a Deep Learning (DL) system for the automated detection and classification of nipple damage. Using a dataset of 1,090 images from clinical trials developed in S\u0026atilde;o Paulo, Brazil, we implemented a Resnet50 convolutional neural network (CNN) to perform two tasks: (1) binary classification to differentiate between intact nipples and those with damages and (2) multiclass classification to identify four types of damage (closed wound, crust, erosion, and fissure) based on the instrument for classifying nipple and areola complex lesions. Data augmentation techniques were applied to upsample the dataset to 8,720 images. The binary classification model achieved an average area under the receiver operating characteristics curve (AUROC) of 0.99 and a recall of 95.90%, demonstrating high accuracy in detecting nipple damage. The multiclass model achieved AUROC values ranging from 0.89 to 0.99 in nipple damage classification, with the highest performance observed for closed wounds (AUROC\u0026thinsp;=\u0026thinsp;0.98) and erosion (AUROC\u0026thinsp;=\u0026thinsp;0.99). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed the model\u0026rsquo;s focus on damaged areas, which aligned closely with clinical assessments. Our findings highlight the potential of DL to improve lactation care by enabling accurate, automated nipple damage classification, particularly in settings with limited access to lactation specialists. This study represents a significant step toward leveraging technology to address challenges in lactation care and improve outcomes for breastfeeding mothers.\u003c/p\u003e","manuscriptTitle":"Automated Detection and Classification of Nipple Damage in Lactation Care","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-08 12:28:45","doi":"10.21203/rs.3.rs-6348223/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-08-30T00:10:05+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-09T13:26:03+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"84882479714418075866918825370367975885","date":"2025-07-18T13:40:21+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-16T16:24:22+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"263000195104182275062006723151094828428","date":"2025-04-26T08:24:40+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"42687276368651273335275762868971228270","date":"2025-04-26T05:24:02+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-23T23:31:37+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-03T19:16:26+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-02T14:49:31+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Women's Health","date":"2025-04-01T00:24:18+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-womens-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [npj Women's Health](https://www.nature.com/npjwomenshealth/)","snPcode":"44294","submissionUrl":"https://submission.springernature.com/new-submission/44294/3","title":"npj Women's Health","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"652e0b6d-c822-41d3-aa81-f5a9407832e6","owner":[],"postedDate":"April 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":46655962,"name":"Health sciences/Health care"},{"id":46655963,"name":"Health sciences/Health care/Diagnosis"}],"tags":[],"updatedAt":"2025-08-30T00:23:06+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-08 12:28:45","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6348223","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6348223","identity":"rs-6348223","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}