{"paper_id":"1630291c-2f31-4eab-bd25-6a23d2214bcc","body_text":"(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nEndometriosis Lesion Classification Using Deep\nTransfer Learning Techniques\nShujaat Ali Zaidi 1, Varin Chouvatut2*, Chailert Phongnarisorn 3\nDepartment of Computer Science-Faculty of Science, Chiang Mai University, Mueang, Chiang Mai 50200, Thailand 1,2\nDepartment of Obstetrics and Gynecology-Faculty of Medicine, Chiang Mai University,\nMueang, Chiang Mai 50200, Thailand 3\nAbstract—In resource-limited settings, assisting physicians\nwith disease identification can significantly improve patient\noutcomes. Early diagnosis is crucial, as many patients could\nremain healthy with timely intervention. Recent advancements\nin deep learning models for medical image processing have\nenabled algorithms to achieve diagnostic accuracy comparable\nto that of healthcare professionals. This research aims to develop\na comprehensive system for the rapid and precise detection\nof endometriosis lesions. We explore the several deep transfer\nlearning architectures, specifically MobileNetV2, VGG19, and\nInceptionV3, on the Gynecologic Laparoscopy Endometriosis\nDataset (GLENDA). Through extensive literature review and\nparameter optimization, we find that MobileNetV2 outperforms\nthe other models in terms of accuracy. However, challenges\nremain, as healthcare imaging datasets often suffer from limited\nsample sizes and uneven class distributions. Collecting additional\nsamples can be costly and time-consuming, which is a prevalent\nissue in medical imaging. To address this, we employ Deep Convo-\nlutional Generative Adversarial Networks (DCGAN) to enhance\nthe dataset by generating synthetic images, thus improving class\nbalance. This image augmentation strategy not only boosts model\nperformance but also reduces the manual effort required for\nimage labeling. We evaluate our proposed model using metrics\nsuch as accuracy, precision, recall, and F1-score. Initially, our\nmodel achieves an accuracy of 95%. The introduction of synthetic\nsamples results in an increased accuracy of 99%, reflecting a 4%\nimprovement and enhancing the model’s overall efficacy.\nKeywords—Endometriosis classification; lesion detection; med-\nical image classification; deep learning; transfer learning; DCGAN\nI. INTRODUCTION\nEndometriosis is a medical condition in which endometrial\ncells, naturally found within the uterus, begin to grow outside\nthe uterine cavity [1], [2]. The preliminary hypothesis describ-\ning this is retrograde menstruation, where menstrual tissue\npursues a distinctive path. Rather than being expelled during\nthe menstrual cycle, the tissue flows in reversal, traveling\nup through the fallopian tubes and potentially implanting\nin the ovary or within the abdominal cavity, resulting in\nlesion construction. These lesions are differ in size, with\nclassification based on dimensions critical to diagnosis and\ntreatment. Mainly, lesions measuring five millimeters or more\nare typically classified as deep-seated endometriosis [3], [4].\nEndometriosis is frequently associated with significant\ndiagnostic delays [5], [6]. Endometriosis disorder has a\nbroad spectrum of symptoms, such as menstrual pain, acute\n*Corresponding authors.\npelvic pain, infertility, and pain during intercourse. Non-\ngynecological symptoms can also appear, including pain while\nurinating, discomfort during bowel movements, flank pain,\nfatigue, and blood in the urine, among others. Furthermore,\nthe biological exams are complicated, and results can differ\nwidely, even in examinations conducted by professionals in\nthe field [7], [8].\nEndometriosis is a severe gynecological problem impacting\nan estimated 190 million women internationally [9], [10]. The\ndisease is prevalent among women of different age class, from\ninfant women to those who are post-menopause. Studies esti-\nmate that approximately 10 percent of women of reproductive\nage are affected by endometriosis [11], with a prevalence\nof 2 percent to 4 percent among postmenopausal women\n[12]. Additionally, The disease’s influence on young women\nis significant, with nearly 50 percent of those undergoing\npersistent pelvic pain occurring before the age of 20 being\ndiagnosed with endometriosis [4].\nEndometriosis can be diagnosed through various pro-\ncedures, including ultrasound, magnetic resonance imaging\n(MRI), and laparoscopy [13], [14]. Laparoscopy is the gold\nbenchmark, as it permits explicit examination of the abdominal\nand pelvic areas through a camera inserted via a small incision.\nMRI, a non-invasive imaging method, uses magnetic areas to\ncreate detailed images of the body’s internal structures, making\nit beneficial in recognizing deep lesions or more extensive\ncysts associated with endometriosis. Ultrasound, generally per-\nformed transvaginally, is usually utilized for initial screening\nand assists in detecting ovarian cysts or other abnormalities\n[15], [16], [17], [18].\nIn recent years, computer-aided methodologies have be-\ncome essential in medical fields for disease diagnosis, in-\ncluding identifying conditions such as heart disease [19],\nendometriosis lesions detection [20], and classifying breast\ncancer [21]. However, Healthcare imaging datasets frequently\nsuffer from insufficiency and uneven class distributions. More-\nover, acquiring additional samples is both expensive and time-\nintensive. This is a common problem in the medical image do-\nmain. Researchers endeavor to conquer this problem by utiliz-\ning data augmentation. Image augmentation methods typically\nfall into two main categories: conventional approaches and\ndeep learning techniques. Conventional image augmentation\nmethods apply simple manipulations like translation, cropping,\nflipping, or resizing. Deep learning-based augmentation uti-\nlizes state-of-the-art neural networks to create more adaptive,\ndata-rich transformations. These two techniques enhance data\ndiversity, helping models to generalize more effectively [22],\nwww.ijacsa.thesai.org 841|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\n[23].\nA promising deep-learning technique for image synthesis\ninspired by the game hypothesis is the Generative Adversar-\nial Network (GAN) [24], [25]. In GANs, two networks are\nlearned in an adversarial manner: one network, the generator,\nconstructs artificial images, while the other, the discrimina-\ntor, learns to distinguish between real and synthetic images.\nBy repeating this adversarial approach, GANs enhance both\nnetworks’ abilities, allowing the generation of high-quality,\nrealistic images. The computer vision community has adopted\nGANs, directing to various variations developed for photore-\nalistic image generation [26], [27].\nDeep learning has attracted significant attention in many\nfields (for more detail, see [28], [29], [30], [31], [32]) such\nas the medical domain because of its efficacy in image\nanalysis. It is employed for classification, enhancement of\nimage quality, and segmentation of medical images. In recent\nyears, deep learning techniques have rapidly evolved, with\nMobileNetV2 attaining popularity due to its compact archi-\ntecture, which makes it perfect for various applications [33].\nThe MobileNetV2 architecture contains an internal structure\nwith a linear bottleneck, an element that minimizes memory\nrequirements for better processing. Therefore, to provide an\naccurate and efficient classification of endometriosis lesions,\nin this study, we propose an endometriosis image classifica-\ntion founded on the MobileNetV2 architecture. The proposed\nmethod uses MobileNetV2 as the base model for the transfer\nlearning process. We add a global pooling layer and two fully\nconnected layers to enhance the model’s performance and\nrefine classification outcomes [34]. The primary contributions\nof this research are as follows.\n•We propose a synthesis of high-quality endometrio-\nsis lesions from laparoscopy images employing deep\nconvolutional generative adversarial networks (DC-\nGANs).\n•Several state-of-the-art deep learning architectures,\nsuch as VGG19, InceptionV3, and MobileNetV2, are\nanalyzed for this research. These architectures are fine-\ntuned on the laparoscopy images dataset to obtain\nefficient and accurate endometriosis lesion detection.\n•We improve the training set of our deep learning\nmodel by including synthetic data, a technique that\nguides to enhanced classification results.\nThe remainder of the article is organized as follows: Sec-\ntion 2 explains the literature review on utilizing deep learning\nto analyze endometriosis lesions. Sections 3 and 4 present\nthe proposed methodology and experimental result analysis.\nConsequently, Sections 5 and 6 provide the discussion and\nconclusion of the study, respectively.\nII. LITERATUREREVIEW\nDetecting endometriosis through laparoscopy imaging is\nchallenging due to the disease’s complex and varied presen-\ntation. Its lesions usually have subtle or indistinct characteris-\ntics, making them challenging to identify accurately, even by\nspecialists. Machine learning applications aimed at diagnosing\nendometriosis through laparoscopy imaging still need to be\ndeveloped, partially due to the limited access to labeled, rep-\nresentative datasets. This section summarizes recent research\nin machine learning for endometriosis diagnosis, focusing on\ntechniques, methodologies, and the insights they provide.\nZaidi [5] employed a deep-learning based approach to\ndetect endometriosis lesions from laparoscopy images.The\nteam achieved a accuracy of 0.93% by applying Inception V3\nmodel with 5-fold cross-validation techniques. Deep convolu-\ntional networks, exemplified as GoogLeNet [35], have been\neffectively employed across various applications, showcasing\ntheir versatility and robust feature extraction capabilities. These\nnetworks serve as robust backbones for many state-of-the-art\ndeep-learning architectures used in image and video research\nanalysis. By leveraging hierarchical layers of convolutional\nfilters, they can capture complicated patterns and attributes,\nmaking them ideal for object recognition, classification, and\nsegmentation tasks. Visalaxi [36] employed a deep-learning\napproach to categorize laparoscopy images associated with\nendometriosis. They employed the dataset for gynecologic\nlaparoscopy related to endometriosis [37], comprising about\n6,000 images from laparoscopy videos. Sixty percent of this\ndataset was used for training, and several architectures were\ntested to determine the model with the best performance. The\nResNet50 architecture [38], loaded with ImageNet pre-trained\nweights [39], achieved the highest accuracy at 90 percent,\nwith sensitivity and precision scores of 82 percent and 83\npercent, respectively.Leibetseder’s study used transfer learning\nand the Faster R-CNN [40] and Mask R-CNN [41] models to\naccurately detect endometriosis in laparoscopic images, getting\na 32.4 percent precision. In addition, the research investigated\ndifferent data augmentation methods, revealing that a blend of\ncropping and rotation yielded optimal results. The GLENDA\ndataset was employed as the primary data source in this\nresearch.\nIn 2021, Yun [42] and colleagues introduced a neural\nnetwork model specifically designed to aid in classifying\nendometriosis. This study employed a convolutional neural net-\nwork (CNN) architecture called VGGNet-16. Using a dataset\nof 6,478 histopathology images for training. The researchers\naimed to develop a highly accurate system that could sup-\nport and potentially enhance the diagnostic work performed\nby radiologists. In their research, Takahashi and colleagues\n[43] explored how computer vision techniques can support\ndetecting endometriosis cancer in laparoscopic images. The\nteam achieved a prediction accuracy of 90.29% by applying\nadvanced neural network techniques. Sudalaimuthu [44] pro-\nposed an innovative approach recognized as Structural Similar-\nity Analysis of Endometriosis (SSAE). This approach evaluates\nendometriosis progression in areas such as the ovaries, uterus,\nrectum, and peritoneum founded on laparoscopy images from\nthe GLENDA image dataset. To enhance the model’s robust-\nness, images were subjected to data augmentation techniques,\nincluding horizontal and vertical shifts, rotation, shear, and\nzoom. The data was split, with 70 percent allocated for training\npurpose and the 30 percent for testing purpose. The U-Net\narchitecture was utilized to explore factors like filter sizes and\noptimization methods. The highest results recorded were dice\ncoefficient is 0.74 and an IoU is 0.72, though it is unclear\nif these metrics were specific to the test set.The literature on\nautomated identification and segmentation of endometriosis re-\nmains limited, with few works addressing these areas. Among\nwww.ijacsa.thesai.org 842|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nthe existing studies, most rely on images from laparoscopy\nprocedures, which, while invasive, are the gold benchmark for\ndiagnosis. Despite using these high-quality images, automated\napproaches are still being developed with the aim of improving\naccuracy and performance.\nOur research differentiates itself from the abovementioned\nresearch by suggesting a data augmentation approach to\nconventional and deep learning approaches that encompass\ndetection and classification tasks for analyzing deep or deep-\nseated endometriosis. Additionally, it focauses on laparoscopy\nimages, seeking at premature examination beyond the essential\nfor invasive procedures.\nIII. MATERIALS ANDMETHODS\nThis section provides a detailed description of the materials\nused and the approach taken for classifying endometriosis\nlesions, including an overview of the image collection process,\na summary of the overall methodology, and a discussion of the\nproposed methodology. Figure(1) provides a flowchart of the\nmethodology followed in this research.\nFig. 1. A Flowchart of proposed methodology.\nA. Data Collection\nThe Gynecologic Laparoscopy Endometriosis Dataset\n(GLENDA) is a comprehensive dataset derived from more\nthan 400 gynecologic laparoscopy video recordings, a signifi-\ncant number of which depict instances of endometriosis with\nvarying phases of severity. GLENDA comprises over 25,000\nimages, including more than 12,000 positive pathological\nphotographs related to endometriosis and over 13,000 negative\nnon-pathological images devoid of prominent endometriosis.\nThe dataset is intentionally designed for various artificial\ncontent analysis tasks related to endometriosis recognition. The\ndimensions of the images are 640 by 360 pixels.\nB. Image Pre-processing\nImage preprocessing is a collection of procedures used\non raw images to prepare them for additional examination\nor processing. The primary aim is to improve the image\nby identifying pertinent information while reducing artifacts,\nnoise, and extraneous features that might interfere with fur-\nther analysis. Prevalent preprocessing tasks include resizing,\nfiltering, cropping, noise attenuation, color modification, and\nimage enhancement. The phases are chosen according to the\nimage properties and the particular application’s requirements.\nIn this research work, All images were resized to 224×224\npixels. We also applied image sharpening to the entire dataset\nto improve the quality of the images. Furthermore, the images\nare processed in RGB format. We implemented many rescaling\nmethods, including multiplying each pixel by 1/255. This\ngeneralizes the input, conserves memory, and reduces the\ncomputational expense of applicable procedures. Moreover, it\nalso facilitates the understanding of the ideal function.\nC. Data Augmentation Using Image Data Generator Methods\nThis approach is used to generate data examples from the\ncurrent samples. It is beneficial when the dataset contains\na limited number of instances or exhibits class imbalance.\nWe may use many geometric techniques to generate aug-\nmented data. This approach facilitates the learning of a class-\nimbalanced dataset and enhances the model’s generalization,\nreducing overfitting. The model acquires the ability to handle\nunexplored versions of training samples. We used the Image\nData Generator package from Kera’s to do image augmenta-\ntion. The following are the approaches that were executed for\nthis objective:\n1) Random zoom:It dynamically zoomed into the provided\nimage. In this work, the zoomed value is fixed to 0.1 percent.\n2) Random rotation:It dynamically rotates the provided\nimage based on the specified value. The rotated value is fixed\nat 30 degrees.\n3) Horizontal flip:This option randomly inverts the image\nhorizontally to alter the positioning of its sides.\n4) Width shift:It horizontally alters the image’s width, with\na value established at 0.2 percent of its overall width.\n5) Height shift:It vertically alters the image’s height, with\na value established at 0.2 percent of its overall height.\nwww.ijacsa.thesai.org 843|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nD. Data Augmentation Using DCGAN Method\nDeep Convolutional Generative Adversarial Networks (DC-\nGANs) are a class of generative models that utilize deep\nconvolutional layers to generate realistic images from random\nnoise. They are built upon the framework of traditional GANs,\nwhich consist of two neural networks: a generator and a\ndiscriminator. Figure ( 2 and 3) illustrates its architecture.\nThe generator learns to produce synthetic images, while the\ndiscriminator evaluates their authenticity. Both networks are\ntrained simultaneously in an adversarial process, where the\ngenerator aims to create images that can deceive the discrim-\ninator, and the discriminator strives to distinguish between\nreal and generated images.The generator creates progressively\nmore naturalistic pictures throughout the training to deceive the\ndiscriminator. This adversarial strategy incrementally enhances\nthe quality of the produced pictures. The generator learns\nexplicitly to translate random noise vectors into picture spaces\nthat mimic the distribution of authentic endometriosis images.\nAt the same time, the discriminator acquires the ability to\ndifferentiate between actual and faked images. Ultimately,\nthis leads to the generator creating high-fidelity synthetic\nendometriosis pictures that closely mimic authentic ones.\nDCGANs introduce specific architectural enhancements,\nincluding the use of convolutional layers instead of fully\nconnected layers, which improves spatial structure preserva-\ntion in the generated images. Additionally, techniques like\nbatch normalization and the use of Leaky ReLU activation\nfunctions enhance training stability and convergence. These\nimprovements enable DCGANs to generate high-quality and\ncoherent images across various domains, such as medical\npicture classification, and objects.\nThe structured use of convolutional layers helps the model\nlearn hierarchical representations, which are crucial for pro-\nducing realistic outputs. Furthermore, the adversarial training\napproach encourages the generator to refine its outputs itera-\ntively, leading to visually convincing results. Due to their effec-\ntiveness and simplicity, DCGANs have become a foundational\nmodel in the field of generative image synthesis and serve as a\nbasis for more advanced architectures in related applications.\nDCGANs can generate data from existing samples. In sit-\nuations like medical picture classification, gathering sufficient\ndata to build a big deep learning model is crucial, which might\nbe arduous. Utilizing presently accessible data, we may pro-\nduce new samples with DCGANs; these synthetic samplings\ndo not directly correspond to actual patients. This technology\nhas expedited the use of deep learning in medical imaging;\nnonetheless, it requires significant improvement in accurately\ncollecting delicate tissues from images. In this study employs a\ndeep convolutional GAN (DCGAN) [46] to generate synthetic\nendometriosis images featuring pathologies diseases in the\nused dataset. A deep generative model of around five layers\nwas used with a relatively slight discriminator network. The\ncomplex architecture assists in detecting subtle tissues within\nthe endometriosis images. DCGAN was used for this analysis\nbecause of its stability in terms of simplicity and efficacy in\nproducing high-quality pictures. In contrast to more intricate\nGAN variations like CGAN, BEGAN, WGAN, and DCGAN,\nit offers a simple architecture that is easy to build and tune,\nrendering it appropriate for our application in medical image\ncreation. Moreover, DCGAN has been extensively evaluated\nFig. 2. Generator architecture [49].\nFig. 3. Discriminator architecture [49].\nacross many image synthesis tasks, showing its reliability and\neffectiveness in creating realistic images, which coincides with\nour objective of enhancing medical datasets.\nWe trained DCGANs to synthesize images of endometrio-\nsis lesions, with separate models developed for each lesion\ncategory. The training process involved alternating updates\nto the generator and discriminator networks in an iterative\nmanner, allowing both networks to improve their performance\nprogressively. A learning rate of 0.0002 was employed, and\ntraining was conducted for 200 epochs for each lesion category\nto ensure adequate learning of the underlying data distribution.\nThe input dataset consisted of original images with a res-\nolution of 224 × 224 pixels, and a batch size of 128 was used\nduring training to optimize computational efficiency. Upon\ncompletion of the training process, we generated 300 synthetic\nimages for each lesion category. Each class-specific training\nsession required approximately 20 hours to complete on a local\nPC equipped with an Intel i7-7820HQ CPU. These synthesized\nimages were then merged with the original dataset to augment\nthe data available for subsequent analysis or model training.\nThis approach ensured an expanded and balanced dataset that\ncould potentially improve downstream performance in tasks\nsuch as endometriosis lesions classification or segmentation.\nTo evaluate the visual quality of the synthesized images,\na radiologist with expertise in endometriosis assessed the\nsynthetic lesions for their resemblance to real clinical cases.\nThe radiologist’s analysis focused on key morphological fea-\ntures, such as texture, contrast, and spatial distribution, to\ndetermine whether the synthetic lesions accurately mirrored the\nappearance of genuine endometriosis implants. By conducting\ntraining independently for each lesion category and ensuring\nconsistency in the training parameters, we aimed to achieve\nhigh-quality image synthesis across all categories. This method\nwww.ijacsa.thesai.org 844|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nFig. 4. DCGANs synthetic images.\nhighlights the potential of DCGANs in generating synthetic\ndata that closely resembles real-world medical images, con-\ntributing to the development of robust datasets in medical\nimaging research.Figure (4) displays examples created for the\npathology class after 200 epochs.\nE. Training Phase\nThis section emphasizes some of the primary details ap-\nplied in the training stage, such as kinds of models, hyperpa-\nrameters, and classifier architectures.\n1) Dataset split:In supervised machine learning and deep\nlearning workflows, datasets are first divided into distinct\nsets for training, validation, and testing. Image augmentation,\ncommonly performed on the training and potentially validation\nsets, aids in model generalization during the training and op-\ntimization stages. To ensure unbiased results, test data usually\nremains unaugmented to evade data leakage. Nevertheless,\nrecent research highlights that augmenting test data can also\nbe valuable in examining model robustness.In this work, the\ndataset is divided into 70% for training, 20% for validation,\nand 10% for testing.\nFig. 5. MobileNetV2 architecture [47].\n•Training Images: 17977\n•Validation Images: 5136\n•Testing Images: 2569\n2) Model architectures:In this study, several deep learning\nmodels were evaluated using the Glenda dataset to assess\ntheir performance across different architectures. Three mod-\nels (VGG19, InceptionV3, MobileNetV2)were implemented\nutilizing the Keras library in Python. Transfer learning was\nutilized for the convolutional layers to initialize these models,\nwhile random initialization was applied to the fully connected\nlayers. The classifier network was built with a custom design\nthat included two hidden layers and an output layer containing\ntwo neuron units, each assigned to a specific class. This\nminimalist design, featuring fewer layers and neurons, was\nstrategically utilized to avoid overfitting.\n3) MobileNetV2:MobileNetV2 [48] is recognized as a\nlightweight convolutional neural network (CNN) that is gen-\nerally utilized in miscellaneous applications. It improves the\nMobileNetV1 model by introducing new modules, especially\ninverted residuals and linear tie-ups. The essence design of\nMobileNet is established on depthwise divisible convolution.\nThis approach differs from standard 2-dimensional convolu-\ntion, which treats all input channels uniformly to produce one\noutput channel. Instead, depthwise convolution applies filters\nindependently to each input channel, resulting in separate\noutput channels that are subsequently combined. The separable\ndepth-wise convolution process involves a subsequent 1×1\npoint-wise convolution, which merges these output channels\ninto one final channel. This method provides the same output\nas traditional convolution while being more efficient due to a\nreduced number of parameters.Figure 5 presents the graphical\nrepresentation of the MobileNetV2 framework.\n4) Hyper parameters:In this study, we kept hyper param-\neters uniform across all models, allowing for a precise and\nfair comparison of each model’s performance. This consistent\napproach allowed for a clear assessment of each model’s\nrelative performance.\nwww.ijacsa.thesai.org 845|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\na) Optimizer:In our optimization process, we rigor-\nously tested two methods: Stochastic Gradient Descent (SGD)\nand Adaptive Momentum Estimation (Adam). After multiple\niterations, we found that both optimizers performed similarly.\nHowever, we chose the Adam optimizer for training, as it is\na more advanced technique for learning model weights and\nhas been proven to enhance convergence and stability in deep\nlearning models.\nb) Learning rate:Since we utilized pre-trained weights\nthrough transfer learning for the convolutional networks in\nour models, we opted for a low learning rate, set explicitly\nat 0.0001, to ensure stable training.\nc) Activation functions:ReLU was implemented across\nall layers, excluding the output layer, adding non-linearity to\nenhance model performance. In the output layer, we applied\nthe softmax function, which is well-suited for labeled classifi-\ncation, as it allows each neuron to generate separate outputs.\nd) Loss function:We opted for Binary Cross-Entropy\n(BCE) to calculate loss, which allows us to consider each label\nindependently.\nIV. RESULTS\nThis section presents a detailed account of the results\nobtained from our experimental research and describes the\nprogramming environments employed to achieve these results.\nFor clarity, each aspect is discussed in separate subsections.\nA. Experimental Setup\nThe large dataset influenced our choice of programming\nenvironments to conduct deep learning experiments. We car-\nried out model implementation and training on our Dell laptop,\nassembled with a 1.80 GHz Intel Core i5 CPU and 16 giga-\nbytes of RAM. Our approach leveraged Python libraries such\nas TensorFlow, Keras, and sci-kit-learn for effective attribute\nclassification. Jupyter Notebook enabled efficient program\ndevelopment and data analysis, and we observed satisfactory\nresults across several models.\nB. Evaluation Procedure\nIn order to make a comprehensive comparison of model\nclassification performance, this paper leverages various metrics\nassociated with the confusion matrix, including precision,\nrecall, F1-score, and accuracy.Accuracy calculates the model’s\nprevalent correctness, while precision focuses on the reliability\nof optimistic predictions. Recall quantifies the model’s sensi-\ntivity by measuring the proportion of true positives, and speci-\nficity assesses the ability to identify negative cases correctly.\nWe clearly understand each model’s strengths and weaknesses\nby calculating and analyzing these metrics. The formulas for\ncalculating these performance metrics are provided below.\nAccuracy= T P+T N\nT P+T N+F P+F N (1)\nPrecision= T P\nT P+F P (2)\nRecall= T P\nT P+F N (3)\nF1-score= 2· Precision·Recall\nPrecision+Recall (4)\nIn this equation, TP, or True Positive, refers to the count\nof positive cases accurately identified as positive. FN, or False\nNegative, represents the number of positive cases mistakenly\nclassified as negative. FP, or False Positive, denotes the count\nof negative cases incorrectly labeled as positive, while TN, or\nTrue Negative, captures the number of negative cases correctly\nclassified as negative.\nC. Area Under the Curve (AUC)\nThe ROC curve is commonly used to evaluate how well\na model can distinguish between different classes. It is based\non plotting the True Positive Rate (TPR) and False Positive\nRate (FPR) across different thresholds. Since TPR and FPR are\ndirectly proportional, they increase together. The ROC curve\nvisually represents this relationship and enables calculation of\nthe Area Under the Curve (AUC), where a larger AUC value\nimplies a more robust classification capability. Thus, a higher\nAUC indicates better model performance.\nD. VGG19\nIn analyzing endometriosis lesion images, spatial data is\ncrucial for effective disease classification. Deep learning-based\nmodels are frequently employed to capture these essential\nfeatures, and VGG19, a highly layered convolutional model,\nwas chosen for this study. We applied VGG19 with pre-trained\nweights for feature extraction and added a custom three-layer\nclassifier. The model was trained across three configurations\nwith varying epochs (3, 7, 14, and 20) and a batch size of\n(32, 64, and 128), resulting in above-average performance\nmetrics with an accuracy of up to 88%. Table I compiles the\nresults of tests conducted without data augmentation, providing\ninsights into the baseline performance of our endometrio-\nsis lesion detection algorithm. Table II, on the other hand,\npresents the outcomes when data augmentation techniques are\napplied, showcasing the improvements in detection accuracy\nand robustness achieved through augmentation. Figures 6 and 7\nillustrate the loss trajectories and accuracy trends, respectively,\nfor the VGG19 model trained on the baseline dataset (without\naugmentation). In contrast, the effects of incorporating data\naugmentation are evident in Figures 8 and 9, which depict the\nmodel’s loss dynamics and classification performance under\nenhanced training conditions. Figure 10 shows the confusion\nmatrices illustrating the endometriosis lesion classification\noutcomes from testing the VGG19 model. The AUC-ROC\ncurve, which can be seen in Figure 11, also explains the result\nof the VGG19 model.\nE. InceptionV3\nThe InceptionV3 model offers an optimized approach to\nachieve high classification accuracy without relying on an ex-\ntensive network architecture. With fewer learnable parameters\nthan the VGG19 model, InceptionV3 provides an efficient al-\nternative that still delivers excellent classification performance.\nwww.ijacsa.thesai.org 846|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nTABLE I. VGG19 MODELRESULTSWITHOUTDATAAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 56 54 55 56\n64 57 57 56 57\n128 59 59 59 60\n7\n32 63 63 63 63\n64 65 64 64 65\n128 66 65 66 66\n14\n32 70 70 69 70\n64 74 73 73 74\n128 77 77 77 78\n20\n32 80 80 79 80\n64 80 79 80 80\n128 81 81 81 82\nFig. 6. VGG19 loss display without data augmentation.\nFig. 7. VGG19 accuracy display without data augmentation.\nIts design prioritizes streamlined architecture, balancing effec-\ntiveness with computational efficiency.InceptionV3 model was\ntrained across three configurations with varying epochs (3, 7,\n14, and 20) and a batch size of (32, 64, and 128), resulting\nin above-average performance metrics with an accuracy of up\nto 97%, that are higher than VGG19. Table III compiles the\nresults of tests conducted without data augmentation, providing\ninsights into the baseline performance of our endometriosis\nlesion detection algorithm. Table IV, on the other hand,\npresents the outcomes when data augmentation techniques are\nTABLE II. VGG19 MODELRESULTSWITHDATAAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 60 60 61 61\n64 62 61 62 63\n128 64 63 64 65\n7\n32 68 69 69 70\n64 72 72 73 73\n128 74 74 75 75\n14\n32 77 76 77 78\n64 79 79 79 80\n128 82 83 83 84\n20\n32 84 84 85 85\n64 87 87 87 88\n128 88 87 88 88\nFig. 8. VGG19 loss display.\nFig. 9. VGG19 accuracy display.\napplied, showcasing the improvements in detection accuracy\nand robustness achieved through augmentation. Figures 12\nand 13 illustrate the loss trajectories and accuracy trends,\nrespectively, for the InceptionV3 model trained on the baseline\ndataset (without augmentation). In contrast, the effects of\nincorporating data augmentation are evident in Figures 14 and\n15, which depict the model’s loss dynamics and classification\nperformance under enhanced training conditions. Figure 16\nshows the confusion matrices illustrating the endometriosis\nlesion classification outcomes from testing the InceptionV3\nwww.ijacsa.thesai.org 847|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nFig. 10. VGG19 confusion matrix.\nFig. 11. VGG19 AUC-ROC curve.\nmodel. The AUC-ROC curve, which can be seen in Figure\n17, also explains the result of the InceptionV3 model.\nTABLE III. INCEPTIONV3 MODELRESULTSWITHOUTDATA\nAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 65 63 64 65\n64 67 67 67 67\n128 70 69 69 70\n7\n32 73 72 73 73\n64 75 75 75 76\n128 80 79 79 80\n14\n32 83 83 82 83\n64 87 85 86 87\n128 88 88 89 89\n20\n32 90 89 89 90\n64 90 91 90 91\n128 92 91 92 92\nF . MobileNetV2\nMobileNetV2 builds on the foundation of MobileNetV1\nwith an updated architecture featuring fewer learnable param-\neters for greater efficiency.MobileNetV2 model was trained\nacross three configurations with varying epochs (3, 7, 14,\nFig. 12. InceptionV3 loss display without data augmentation.\nFig. 13. InceptionV3 accuracy display without data augmentation.\nTABLE IV. INCEPTIONV3 MODELRESULTSWITHDATAAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 52 53 53 54\n64 56 56 57 58\n128 59 58 58 59\n7\n32 62 63 61 63\n64 68 69 71 72\n128 80 80 82 83\n14\n32 89 88 89 90\n64 90 91 89 92\n128 92 92 92 93\n20\n32 92 93 93 94\n64 95 94 96 96\n128 96 97 97 97\nand 20) and a batch size of (32, 64, and 128), resulting in\nabove-average performance metrics with an accuracy of up to\n99%, which are higher than InceptionV3. Table V compiles the\nresults of tests conducted without data augmentation, providing\ninsights into the baseline performance of our endometriosis\nlesion detection algorithm. Table VI, on the other hand,\npresents the outcomes when data augmentation techniques are\napplied, showcasing the improvements in detection accuracy\nand robustness achieved through augmentation. Figures 18\nand 19 illustrate the loss trajectories and accuracy trends,\nwww.ijacsa.thesai.org 848|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nFig. 14. InceptionV3 loss display.\nFig. 15. InceptionV3 accuracy display.\nFig. 16. InceptionV3 confusion matrix.\nrespectively, for the MobileNetV2 model trained on the base-\nline dataset (without augmentation). In contrast, the effects of\nincorporating data augmentation are evident in Figures 20 and\n21, which depict the model’s loss dynamics and classification\nperformance under enhanced training conditions. Figure 22\nshows the confusion matrices illustrating the endometriosis\nlesion classification outcomes from testing the MobileNetV2\nmodel. The AUC-ROC curve, which can be seen in Figure 23,\nalso explains the result of the MobileNetV2 model.\nFig. 17. InceptionV3 AUC-ROC curve.\nTABLE V. MOBILENETV2 MODELRESULTSWITHOUTDATA\nAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 70 70 69 70\n64 70 70 71 71\n128 72 72 71 73\n7\n32 73 74 74 75\n64 76 77 77 78\n128 79 80 79 80\n14\n32 83 84 83 84\n64 83 84 84 85\n128 85 86 86 87\n20\n32 90 90 89 90\n64 92 91 92 92\n128 93 94 94 95\nFig. 18. MobileNetV2 loss display without data augmentation.\nV. DISCUSSION\nEndometriosis is one of the leading gynecological issues\nfacing women worldwide. Diagnosing and treating this con-\ndition remains complex, particularly in settings with lim-\nited medical resources. Deep learning techniques applied to\nmedical imaging on large datasets have enabled computer\nalgorithms to achieve diagnostic accuracy similar to that of\nhealthcare professionals. In this research, we propose a deep\nlearning-based solution for classifying endometriosis lesions\nwww.ijacsa.thesai.org 849|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nFig. 19. MobileNetV2 accuracy display without data augmentation.\nTABLE VI. MOBILENETV2 MODELRESULTSWITHDATA\nAUGMENTATION\nEpochs Batch-Size Precision (%) Recall (%) F1-Score (%) Accuracy (%)\n3\n32 61 62 61 62\n64 61 61 62 63\n128 65 65 65 65\n7\n32 67 66 67 68\n64 70 70 71 73\n128 79 80 79 80\n14\n32 83 84 82 84\n64 86 87 87 87\n128 91 91 91 92\n20\n32 94 93 94 94\n64 95 96 96 97\n128 99 99 99 99\nFig. 20. MobileNetV2 loss display.\nTABLE VII. OVERALLRESULTS\n# Deep Learning\nModels\nAccuracy without\nData Augmentation (%)\nAccuracy with\nData Augmentation (%)\n1 VGG19 82 88\n2 InceptionV3 92 97\n3 MobileNetV2 95 99\nFig. 21. MobileNetV2 accuracy display.\nFig. 22. MobileNetV2 confusion matrix.\nFig. 23. MobileNetV2 AUC-ROC curve.\nin laparoscopy images, utilizing the MobileNetV2 architec-\nture alongside a neural network classifier. We incorporate\nboth conventional methods and deep learning-based image\naugmentation techniques. Experimental findings indicate that\ndeep learning models are well-suited to accurately classify\nendometriosis lesions, facilitating a robust diagnostic tool for\nendometriosis.\nThe incorporation of synthetic images into the training\nprocess offers a substantial benefit to the performance of\ndeep learning models, particularly when dealing with small or\nimbalanced datasets. In many real-world applications, such as\nwww.ijacsa.thesai.org 850|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\nTABLE VIII. COMPAREDWITHDIFFERENTTECHNIQUESEMPLOYED BY\nOTHERRESEARCHERS\n# Citation Classifier Accuracy (%)\n1 Visalaxi et al. (2021) [36] ResNet50 91\n2 Yun et al. (2021) [43] VGGNet-16 90.80\n3 Takahashi al. (2021) [44] DNN 90.29\n4 Leibetseder al. (2022) [40] Faster-RCNN 32.4 (precision)\n5 Sudalaimuthu al. (2022) [45] U-Net 74 (F1-score)\n6 Figueredo al. (2024) [46] Ensemble of Networks 96.67\n7 Zaidi al. (2025) [5] Inception V3 93\n8 Our proposed MobileNetV2 99\nmedical imaging, obtaining a large and representative dataset\ncan be a significant challenge. Synthetic image generation\naddresses this limitation by augmenting the dataset, providing\nthe model with a broader spectrum of examples, including rare\nor underrepresented cases. This allows the model to learn more\nrobust and generalized features, which directly contributes to\nimprovements in key performance metrics such as accuracy,\nprecision, and recall.When compared to models trained solely\non real images, those trained with synthetic images exhibit\nsuperior handling of rare classes. This is particularly important\nin fields like medical imaging, where certain conditions or\nabnormalities may be infrequent but critical for diagnosis. By\nincreasing the diversity of the training data, synthetic images\nhelp to reduce bias toward more common classes, ensuring\nthat the model does not develop a skewed understanding of\nthe data. As a result, models trained with synthetic images are\nbetter equipped to identify and classify rare conditions, leading\nto a decrease in false negatives and an overall improvement in\nrecall.\nMoreover, synthetic images enhance the generalization\nability of deep learning models. With more varied and com-\nprehensive data, models are better prepared to perform well\non unseen data, a crucial aspect for deployment in real-world\nscenarios. This greater ability to generalize ensures that the\nmodel’s performance remains consistent when faced with new,\npreviously unobserved examples. Thus, synthetic images play\na vital role in improving not only the accuracy and reliability of\ndeep learning models but also their ability to adapt to diverse\nand unpredictable data, particularly in specialized fields such\nas medical imaging.\nIn our comparative evaluation of convolutional neural net-\nwork architectures, MobileNetV2 demonstrated superior per-\nformance relative to models such as VGG19 and InceptionV3,\nmainly due to MobileNetV2’s lightweight and efficient struc-\nture, making it ideal for mobile and embedded deployment.\nWith fewer parameters, it offers faster and simpler training\nand deployment. Experimenting with 32, 64, and 128 batch\nsizes, we achieved a favorable trade-off between accuracy and\ncomputational efficiency at a resolution of 224x224. Scaling\npixel values from 0 to 1 was the most effective for normaliza-\ntion. Additionally, we observed that while fine-tuning the fully\nconnected layer and freezing convolutional layers maintained\nstable performance, it did extend convergence time.\nTable VII presents a comparative analysis of various deep\nlearning algorithms tested on the GLENDA dataset, highlight-\ning the performance of each model in detecting endometriosis.\nThis comparison underscores the strengths and limitations of\ndifferent algorithms within the same dataset, allowing for\nan evaluation of each model’s effectiveness in classifying\nendometriosis lesions. Table VIII provides a summary of these\nfindings and compares the performance of our models to\nexisting state-of-the-art approaches, offering insights into how\nour methods advance current standards in accuracy, precision,\nand overall reliability for clinical decision support.\nVI. CONCLUSION\nOur approach, combining the MobileNetV2 model with\nboth conventional and deep learning-based augmentation tech-\nniques, demonstrated high performance, achieving 99% in\nrecall, binary accuracy, and F1-score. The integration of syn-\nthetic samples generated by DCGAN significantly improved\ntraining data diversity and addressed class imbalance issues\nwithin the dataset. Specifically, by generating synthetic images\nof endometriosis using DCGAN, we were able to enhance\nthe MobileNetV2 model’s accuracy, likely due to DCGAN’s\ncapacity to capture diverse manifestations of endometriosis\npresent in the GLENDA dataset. These findings underscore\nthe potential of deep learning models to accurately classify\nendometriosis lesions from laparoscopy images, supporting\ntheir use as clinical decision support tools for timely diagno-\nsis. Additionally, synthetic data augmentation shows promise\nfor addressing similar challenges in other areas of medical\nimaging, offering a pathway to more effective and reliable\ndiagnostic support systems. Future work will aim to further\nimprove system performance by expanding the dataset to\ninclude a wider variety of laparoscopy videos, enhancing both\nmodel accuracy and generalization across clinical scenarios.\nAUTHOR’SCONTRIBUTIONS\nAll authors have accepted responsibility for the entire\ncontent of this manuscript and approved its submission.\nDATAAVAILABILITY\nData availability is not applicable to this article as no new\ndata were created or analysed in this study.\nCONFLICT OFINTEREST\nThe authors state that they do not have any conflicts of\ninterest.\nACKNOWLEDGMENTS\nThe first author is a PhD degree student in the Computer\nScience Program at the Faculty of Science, Chiang Mai\nUniversity (CMU), under the CMU Presidential Scholarship.\nREFERENCES\n[1] J. N. Samreen et al., ‘MRI of endometriosis: a comprehensive review’,\nAppl Radiol, vol. 48, no. 5, pp. 6–12, 2019.\n[2] A. L. Shafrir et al., ‘Risk for and consequences of endometriosis: a crit-\nical epidemiologic review’, Best practice & research Clinical obstetrics\n& gynaecology, vol. 51, pp. 1–15, 2018.\n[3] N. Agarwal and A. Subramanian, ‘Endometriosis-morphology, clinical\npresentations and molecular pathology’, Journal of laboratory physicians,\nvol. 2, no. 01, pp. 001–009, 2010.\nwww.ijacsa.thesai.org 851|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\n[4] H. Kwok et al., ‘Lesion distribution characteristics of deep infiltrating\nendometriosis with ovarian endometrioma: an observational clinical\nstudy’, BMC women’s health, vol. 20, pp. 1–7, 2020.\n[5] S. A. Zaidi, V . Chouvatut, C. Phongnarisorn, and D. Praserttitipong,\n‘Deep learning based detection of endometriosis lesions in laparoscopic\nimages with 5-fold cross-validation’, Intelligence-Based Medicine, vol.\n11, p. 100230, 2025.\n[6] C. M. Becker et al., ‘ESHRE guideline: endometriosis’, Human repro-\nduction open, vol. 2022, no. 2, p. hoac009, 2022.\n[7] Y . Wang, K. Nicholes, and I. M. Shih, ‘The origin and pathogenesis of\nendometriosis’, Annual Review of Pathology: Mechanisms of Disease,\nvol. 15, no. 1, pp. 71–95, 2020.\n[8] E. Pascoal et al., ‘Strengths and limitations of diagnostic tools for\nendometriosis and relevance in diagnostic test accuracy research’, Ul-\ntrasound in Obstetrics & Gynecology, vol. 60, no. 3, pp. 309–327, 2022.\n[9] K. E. Nnoaham et al., ‘Impact of endometriosis on quality of life and\nwork productivity: a multicenter study across ten countries’, Fertility and\nsterility, vol. 96, no. 2, pp. 366–373, 2011.\n[10] C. M. Becker, S. A. Missmer, and K. T. Zondervan, ‘Endometriosis:\nthe authors reply’, New England Journal of Medicine, vol. 383, no. 2,\n2020.\n[11] B. Smolarz, K. Szyłło, and H. Romanowicz, ‘Endometriosis: epidemi-\nology, classification, pathogenesis, treatment and genetics (review of\nliterature)’, International journal of molecular sciences, vol. 22, no. 19,\np. 10554, 2021.\n[12] A. G. Cope, W. M. Vanburen, and S. P. Sheedy, ‘Endometriosis in\nthe postmenopausal female: clinical presentation, imaging features, and\nmanagement’, Abdominal Radiology, vol. 45, no. 6, pp. 1790–1799,\n2020.\n[13] W. K. R. Figueredo, A. C. Silva, A. C. De Paiva, J. O. B. Diniz,\nA. Brand ˜ao, and M. A.P. Oliveira, ‘Automatic segmentation of deep\nendometriosis in the rectosigmoid using deep learning’, Image and Vision\nComputing, 2024.\n[14] S. L. Vlek et al., ‘Laparoscopic imaging techniques in endometriosis\ntherapy: a systematic review’, Journal of Minimally Invasive Gynecology,\nvol. 23, no. 6, pp. 886–892, 2016.\n[15] M. Al-Taher et al., ‘Intraoperative enhanced imaging for detection of\nendometriosis: A systematic review of the literature’, European Journal\nof Obstetrics & Gynecology and Reproductive Biology, vol. 224, pp.\n108–116, 2018.\n[16] C. Mosbrucker, A. Somani, and J. Dulemba, ‘Visualization of\nendometriosis: comparative study of 3-dimensional robotic and 2-\ndimensional laparoscopic endoscopes’, Journal of Robotic Surgery, vol.\n12, no. 1, pp. 59–66, 2018.\n[17] G. Vizzielli et al., ‘Real three-dimensional approach vs two-dimensional\ncamera with and without real-time near-infrared imaging with indo-\ncyanine green for detection of endometriosis: A case-control study’,\nActa Obstetricia et Gynecologica Scandinavica, vol. 99, no. 10, pp.\n1330–1338, 2020.\n[18] J. Quesada et al., ‘Endometriosis: a multimodal imaging review’,\nEuropean journal of radiology, vol. 158, p. 110610, 2023.\n[19] B. Cherradi, O. Terrada, A. Ouhmida, S. Hamida, A. Raihani, and\nO. Bouattane, ‘Computer-aided diagnosis system for early prediction\nof atherosclerosis using machine learning and K-fold cross-validation’,\nin 2021 international congress of advanced technology and engineering\n(ICOTEN), 2021, pp. 1–9.\n[20] C. Nifora, L. Chasapi, M. K. Chasapi, and C. Koutsojannis, Deep\nLearning Improves Accuracy of Laparoscopic Imaging Classification for\nEndometriosis Diagnosis. 2023.\n[21] M. M. Rahman, Y . Ghasemi, E. Suley, Y . Zhou, S. Wang, and J.\nRogers, ‘Machine learning based computer aided diagnosis of breast\ncancer utilizing anthropometric and clinical features’, Irbm, vol. 42, no.\n4, pp. 215–226, 2021.\n[22] H. Ding, N. Huang, and X. Cui, ‘Leveraging GANs data augmentation\nfor imbalanced medical image classification’, Applied Soft Computing,\nvol. 165, 2024.\n[23] M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger, and\nH. Greenspan, ‘GAN-based synthetic medical image augmentation for\nincreased CNN performance in liver lesion classification’, Neurocom-\nputing, vol. 321, pp. 321–331, 2018.\n[24] I. Goodfellow et al., ‘Generative adversarial nets’, in Advances in neural\ninformation processing systems, 2014.\n[25] H. Ding, Y . Sun, N. Huang, Z. Shen, and X. Cui, ‘TMG-GAN:\nGenerative adversarial networks-based imbalanced learning for network\nintrusion detection’, IEEE Transactions on Information Forensics and\nSecurity, vol. 19, pp. 1156–1167, 2023.\n[26] J. Gui, Z. Sun, Y . Wen, D. Tao, and J. Ye, ‘A review on generative adver-\nsarial networks: Algorithms, theory, and applications’, IEEE transactions\non knowledge and data engineering, vol. 35, no. 4, pp. 3313–3332, 2021.\n[27] W. Li et al., ‘EID-GAN: Generative adversarial nets for extremely\nimbalanced data augmentation’, IEEE Transactions on Industrial Infor-\nmatics, vol. 19, no. 3, pp. 3208–3218, 2022.\n[28] A. Turab, F. Ullah, and S. A. Zaidi, ‘A Federated Learning Framework\nfor Intermittent Clients in Collaborative Smart Healthcare Systems’, in AI\nand Blockchain Applications for Privacy and Security in Smart Medical\nSystems, IGI Global Scientific Publishing, 2025, pp. 1–30.\n[29] S. A. Zaidi and V . Chouvatut, ‘Mae Mai Muay Thai Layered Clas-\nsification Using CNN and LSTM Models’, in 2022 26th International\nComputer Science and Engineering Conference (ICSEC), 2022, pp.\n351–356.\n[30] A. Turab, R. Shafqat, S. Muhammad, M. Shuaib, M. F. Khan, and M.\nKamal, ‘Predictive modeling of hepatitis B viral dynamics: a caputo\nderivative-based approach using artificial neural networks’, Scientific\nReports, vol. 14, no. 1, p. 21853, 2024.\n[31] A. Turab, F. Ullah, and S. A. Zaidi, ‘Explainable AI and Computational\nModeling for Real-Time COVID-19 Detection Using Medical Imaging’,\nin AI and Blockchain Applications for Privacy and Security in Smart\nMedical Systems, IGI Global Scientific Publishing, 2025, pp. 161–190.\n[32] S. A. Zaidi and V . Chouvatut, ‘Mae Mai Muay Thai Style Classification\nin Movement Appling Long-Term Recurrent Convolution Networks’,\nJournal of Internet Services and Information Security, vol. 13, no. 1,\npp. 95–112, 2023.\n[33] K. Dong, C. Zhou, Y . Ruan, and Y . Li, ‘MobileNetV2 model for\nimage classification’, in 2020 2nd International Conference on Infor-\nmation Technology and Computer Application (ITCA), IEEE, 2020, pp.\n476–480.\n[34] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. C. Chen,\n‘Mobilenetv2: Inverted residuals and linear bottlenecks’, in Proceedings\nof the IEEE conference on computer vision and pattern recognition, 2018,\npp. 4510–4520.\n[35] P. Ballester and R. Araujo, ‘On the performance of GoogLeNet and\nAlexNet applied to sketches’, Proceedings of the AAAI conference on\nartificial intelligence, vol. 30, 2016.\n[36] S. Visalaxi and T. S. Muthu, ‘Automated prediction of endometriosis\nusing deep learning’, International Journal of Nonlinear Analysis and\nApplications, vol. 12, no. 2, pp. 2403–2416, 2021.\n[37] A. Leibetseder, S. Kletz, K. Schoeffmann, S. Keckstein, and J. Keck-\nstein, ‘GLENDA: gynecologic laparoscopy endometriosis dataset’, in\nInternational Conference on Multimedia Modeling, Cham: Springer\nInternational Publishing, 2019, pp. 439–450.\n[38] K. He, X. Zhang, S. Ren, and J. Sun, ‘Deep residual learning for image\nrecognition’, in Proceedings of the IEEE conference on computer vision\nand pattern recognition, 2016, pp. 770–778.\n[39] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, ‘Imagenet:\nA large-scale hierarchical image database’, in 2009 IEEE conference on\ncomputer vision and pattern recognition, 2009, pp. 248–255.\n[40] A. Leibetseder, K. Schoeffmann, J. Keckstein, and S. Keckstein,\n‘Endometriosis detection and localization in laparoscopic gynecology’,\nMultimedia Tools and Applications, vol. 81, no. 5, pp. 6191–6215, 2022.\n[41] S. Ren, K. He, R. Girshick, and J. Sun, ‘Faster R-CNN: Towards real-\ntime object detection with region proposal networks’, IEEE transactions\non pattern analysis and machine intelligence, vol. 39, pp. 1137–1149,\n2016.\n[42] K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, ‘Mask r-cnn’, in\nProceedings of the IEEE international conference on computer vision,\n2017, pp. 2961–2969.\n[43] Y . Zhang et al., ‘Deep learning model for classifying endometrial\nlesions’, Journal of Translational Medicine, vol. 19, pp. 1–13, 2021.\nwww.ijacsa.thesai.org 852|P a g e\n\n(IJACSA) International Journal of Advanced Computer Science and Applications,\nVol. 16, No. 6, 2025\n[44] Y . Takahashi et al., ‘Automated system for diagnosing endometrial\ncancer by adopting deep-learning technology in hysteroscopy’, PLoS\nOne, vol. 16, no. 3, p. e0248526, 2021.\n[45] S. Visalaxi and T. Sudalaimuthu, ‘Automated segmentation of en-\ndometriosis using transfer learning technique’, F1000Research, vol. 11,\np. 360, 2022.\n[46] W. K. R. Figueredo, A. C. Silva, A. C. de Paiva, J. O. B. Diniz,\nA. Brand ˜ao, and M. A. P. Oliveira, ‘Automatic segmentation of deep\nendometriosis in the rectosigmoid using deep learning’, Image and Vision\nComputing, vol. 151, p. 105261, 2024.\n[47] A. Radford, L. Metz, and S. Chintala, ‘Unsupervised representation\nlearning with deep convolutional generative adversarial networks’, arXiv\npreprint arXiv:1511. 06434, 2015.\n[48] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen,\n‘Mobilenetv2: Inverted residuals and linear bottlenecks’, in Proceedings\nof the IEEE conference on computer vision and pattern recognition, 2018,\npp. 4510–4520.\n[49] N. Zhao, Y . Song, A. Yang, K. Lv, H. Jiang, and C. Dong, ‘Accurate\nClassification of Tunnel Lining Cracks Using Lightweight ShuffleNetV2-\n1.0-SE Model with DCGAN-Based Data Augmentation and Transfer\nLearning’, Applied Sciences, vol. 14, no. 10, p. 4142, 2024.\nwww.ijacsa.thesai.org 853|P a g e","source_license":"CC0","license_restricted":false}