Skin Cancer Classification from Dermatoscopic Images Using Deep Learning Techniques

Skin Cancer Classification from Dermatoscopic Images Using Deep Learning Techniques | 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 Research Article Skin Cancer Classification from Dermatoscopic Images Using Deep Learning Techniques Sukh Singh Oberoi, Ishan Sandhir, Rajalakshmi M. This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6473686/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Skin cancer affects a large number of people and can become serious if not detected early. In many places, timely diagnosis is challenging due to a lack of specialists. Also, skin lesions often look very similar, which makes visual diagnosis tricky. This study looks at using deep learning to help with that. We trained a convolutional neural network model to classify several types of skin cancer using images from the ISIC dataset. Before training, the images were pre-processed and augmented to improve model performance. The model was tested using common metrics like accuracy, precision, recall, and ROC-AUC. The results indicate that AI can support doctors by making skin cancer detection faster and more reliable. Skin Disease Detection Deep Learning Dermoscopic Images AI in Healthcare CNN Classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 I. INTRDOUCTION Skin cancer impacts a significant population and can escalate in severity if not identified promptly. Many of these cases are associated with UV radiation exposure, especially in fair skinned people [1]. Melanoma is known to be the most dangerous of all the types. It can spread quickly and cause serious complications or death unless it is found in time [1]. That’s why early detection is so important. However, in many parts of the world, especially rural or low resource areas, people are not always able to access dermatologists or screening tools. Diagnosis in most cases is based on visual inspection with tools such as dermoscopy, which is not always reliable. The experience of the doctor also plays a very big role in it and that can lead to inconsistency [2]. Artificial intelligence has begun to have a more prominent role in healthcare in recent years. In particular, deep learning has strong results in medical imaging. CNNs are very popular because they can learn “patterns directly from image data without the need for manual feature selection. Doctors can use these models to make faster” and more consistent decisions when they are viewing skin lesions [3]. In this study, we explore a CNN based approach to classify various types “of skin cancer from dermatoscopic images from the ISIC dataset.” Preprocessing and augmentation techniques are used to train the model so that it can generalize better. We aim to develop a system that can be used for early diagnosis and is applicable in places where expert resources are scarce. II. LITERATURE REVIEW Many researchers have looked into using deep learning and AI for identifying skin cancer. Dildar et al. [1] provided an overview of different neural network models for this purpose. They explained that “deep learning models can improve diagnosis if trained with diverse and high-quality datasets” [1]. However, they also mentioned issues with real-time use due to “high computational demands and limited dataset diversity” [1]. Mazhar et al. [2] developed a model that combines CNN and SVM techniques to improve classification. They wrote that “SVM when combined with CNN showed better generalization performance” [2]. But their method had problems detecting rare diseases and produced more false positives. In another study, Esteva et al. [3] used popular deep neural networks like VGG16 and ResNet to identify skin cancer. They found that “deep neural networks achieved dermatologist-level accuracy” [3]. Still, they admitted the system was less reliable when tested with images of people from different skin tones. Claret et al. [4] improved detection by applying wavelet transformation before feeding data into CNNs. According to them, this helped the model pick up useful features. They mentioned, however, that their approach “faced overfitting issues due to small sample sizes” [4]. Mahmud et al. [5] proposed a model that used attention layers to make the predictions easier to understand. They said the attention mechanism “increased transparency of the system” [5]. But like others, their model still needed powerful hardware to run effectively. Recent research has also focused on making skin cancer detection more practical and privacy-friendly. Some teams explored federated learning to avoid sharing patient data [6], while others added attention blocks to better focus on lesions [7]. Akter et al. [8] created a CNN-based model that could classify multiple types of skin cancer and said it “achieved competitive performance across different classes” [8]. III. PROPOSED METHODOLOGY The classification pipeline consists of several stages: data preprocessing, augmentation, CNN-based model construction, training, and evaluation. A custom convolutional neural network is developed to tell between malignant and benign skin lesions. Each stage is designed to optimize performance while minimizing overfitting, given the class imbalance and limited data variability. A. Dataset Description This study uses data from the International Skin Imaging Collaboration (ISIC) archive, a widely recognized benchmark for skin lesion analysis. It comprises high-resolution dermoscopic images annotated and verified by medical experts. For this work, only two broad categories are considered: benign and malignant. Within these categories, the dataset includes a variety of skin lesion types: Benign: Melanocytic nevi, seborrheic keratoses, and dermatofibromas Malignant: Melanomas, basal cell carcinomas, and actinic keratoses These subtypes introduce substantial visual diversity, making the classification task more challenging and realistic. The dataset also provides extra details like type of lesion, diagnosis, and patient details. However, in this study, only the image data and labels were used. B. Data Preprocessing and Augmentation Prior to training, all images were standardized to ensure consistency in model input. Each image was resized to 224×224 pixels, a common input dimension for convolutional neural networks, and pixel intensities were normalized between 0 and 1 to improve numerical stability during training. To address class imbalance—where benign lesions significantly outnumber malignant ones—the dataset was balanced by undersampling the benign class. This ensures that the model gets a balanced set of samples from both classes, reducing bias in predictions and improving sensitivity to malignant cases. To improve generalization and reduce overfitting, data augmentation was employed on the training set. Augmentation techniques included: Random rotations Horizontal and vertical flipping Zooming Width and height shifts These transformations were applied in real-time during training using data generators. By exposing the model to varied representations of the same image, augmentation enhances its ability to learn invariant and robust features, which is critical in medical imaging tasks with limited datasets. C. Model Architecture The proposed model leverages transfer learning by utilizing DenseNet201, “a convolutional neural network pre-trained on the ImageNet dataset, to extract features” [9, p. 4700]. As Huang et al. describe, “each layer obtains additional inputs from all preceding layers and passes on its own feature-maps to all subsequent layers” [9, p. 4700]. This dense connectivity “encourages feature reuse and substantially reduces the number of parameters” [9, p. 4700] compared to traditional CNNs. The model uses a DenseNet201 backbone pre-trained on ImageNet as a feature extractor, followed by a small custom classifier consisting of a Dropout layer, a Dense layer with 512 units, and a Softmax output layer with 9 classes. This structure balances computational efficiency with strong performance on the skin cancer classification task. A full architecture summary is provided in Table I. Table 1. Model Architecture Summary Layer Type Output Shape Parameters Activation DenseNet201 (pretrained) (2, 3, 1920) 18,321,984 ReLU Flatten (11520) 0 - Dropout (0.5) (11520) 0 - Dense (512 units) (512) 5,898,752 ReLU Dense (9 units) (9) 4,617 Softmax Total params: 24,225,353 (92.41 MB) Trainable params: 23,996,297 (91.54 MB) Non-trainable params: 229,056 (894.75 KB) D. Training Configuration “The model was compiled using the Stochastic Gradient Descent (SGD) optimizer with a learning rate of 0.001 and a momentum of 0.9, which was found to provide stable convergence during experimentation. The categorical cross-entropy loss function was used due to the multi-class nature of the classification task. Training was done for 50 epochs with a batch size of 32, and the data was split into 80% for training and 20% for validation. To support model convergence, a learning rate reduction strategy was also applied by monitoring validation accuracy, reducing the learning rate by a factor of 0.5 with a patience of 3 epochs, and setting a minimum learning rate threshold of 0.00001. Training was conducted using a single NVIDIA T4 GPU, and the model required approximately 45 minutes, depending on system load. As shown in Figures 1 and 2, the model demonstrates a steady increase in training and validation accuracy along with a consistent decrease in loss, indicating effective training convergence. E. Evaluation Metrics We evaluated the model using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix metrics, as detailed in Section IV. IV. RESULT AND DISCUSSION A. Result and Evaluation of Model This section presents the experimental results obtained from training and evaluating the proposed model on the skin cancer classification task. Multiple performance metrics are analyzed to assess overall accuracy, class-wise behavior, and model reliability, especially under the medical constraints of misdiagnosis and class imbalance. 1) Accuracy : Used during training and validation to monitor overall performance. “The model achieved a training accuracy of 99.41%, a validation accuracy of 82.40%, and a final test accuracy of 84.00%.” Accuracy curves across epochs are shown in Fig. 2. 2) Confusion Matrix : A 9×9 “confusion matrix was generated to” visualize class-wise prediction performance. This helped identify misclassifications between similar skin cancer types. (Fig. 3) 3) Precision, Recall, F1-Score : A classification report was computed for all nine classes. These metrics are crucial for evaluating performance in tasks with imbalanced classes, particularly in medical diagnostics. As described by Fawcett, "Precision measures the proportion of positive predictions that are actually correct, recall indicates the proportion of actual positives that were correctly identified, and the F1-score is the harmonic mean of precision and recall" [11, p. 862]. 4) Per-Class Accuracy : Individual class-wise accuracy values were also computed to analyze how well the model performs across specific skin cancer types. This is crucial in medical applications where certain types may be harder to detect. (Fig. 5) 5) “ROC (Receiver Operating Characteristic) Curve : ROC” curves were plotted for each class by computing one-vs-rest scores (Fig. 6). As explained in the literature, “the area under the ROC curve (AUC) provides an aggregate measure of performance across all possible classification thresholds” [11]. 6) Precision–Recall (PR) Curve : PR curves were plotted to evaluate performance under class imbalance. These are especially useful when false positives carry lower risk than false negatives, as is often the case in medical screening tasks (Fig. 7). 7) Learning Rate Scheduling : Keras’ ReduceLROnPlateau callback was employed which “Reduce learning rate when a metric has stopped improving” [10] as “models often benefit from reducing the learning rate by a factor of 2–10 once learning stagnates” [10]. B. Discussion “The model demonstrates strong generalization with a high training accuracy of 99.41% and a reasonably good test accuracy of 84.00%. However, the performance gap between training and validation suggests potential overfitting, which could be mitigated further with more data or stronger regularization. Certain classes, such as melanoma and” basal cell carcinoma, showed lower precision and recall, highlighting the challenge of distinguishing between visually similar lesion types. Future work could explore ensemble models, better augmentation, or transfer learning from larger dermatology datasets to improve robustness. V. CONCLUSION This study presents a deep learning-based approach to classify multiple types of skin cancer using dermatoscopic images. The model was trained and tested on the ISIC dataset, which includes nine skin lesion categories. Through extensive preprocessing, data augmentation, and a well-designed convolutional neural network, the model reached a final test accuracy of 84.00%, with promising class-wise performance across most categories. Evaluation using confusion matrix, ROC and precision–recall curves, and class-specific metrics such as F1-score confirmed the model's effectiveness, while also highlighting the challenges in distinguishing visually similar lesion types. The results show that deep learning has strong potential in aiding early detection of skin cancer, particularly as a diagnostic support tool for dermatologists.” VI. FUTURE WORK Although the proposed model shows promising results, several enhancements can be made to improve its performance, accuracy, reliability and real-world applicability: A. Larger and More Diverse Datasets : Incorporating more varied and real-world clinical images, including higher-resolution samples and other demographics, can improve generalization. B. Transfer Learning and Model Ensembling : Using pretrained medical imaging models or combining multiple architectures may enhance feature extraction and robustness. C. Explainable AI (XAI) : Integrating model interpretability techniques (e.g., Grad-CAM or SHAP) can help clinicians trust the model's predictions and understand decision boundaries. D. Mobile/Edge Deployment : Optimizing the model for low-latency edge devices could make it suitable for use in telemedicine and rural healthcare settings. E. Clinical Validation : Collaborating with dermatologists to validate the model in real-world clinical workflows would be a critical next step toward deployment. Declarations Author Contribution S.O. and I.S. wrote the main manuscript text and S.O. added and reviewed citations and references. I.S. prepared introduction. R.M. guided us throughout the project and reviewed the manuscript. Acknowledgments “We sincerely thank the faculty of the Department of Computing Technologies at SRM Institute of Science and Technology for their continuous support and insightful feedback during the course of this project. We are especially grateful to Dr. Rajalakshmi M, Assistant Professor, for her valuable guidance and encouragement throughout the research.” References M. Dildar, S. Akram, M. Irfan, et al., "Skin Cancer Detection: A Review Using Deep Learning Techniques," Int. J. Environ. Res. Public Health, vol. 18, no. 10, p. 5479, May 2021, doi: 10.3390/ijerph18105479. T. Mazhar, I. Haq, A. Ditta, et al ., "The Role of Machine Learning and Deep Learning Approaches for the Detection of Skin Cancer," Healthcare , vol. 11, no. 3, p. 415, 2023, doi: 10.3390/healthcare11030415. A. Esteva, B. Kuprel, R. A. Novoa, et al ., "Dermatologist-level classification of skin cancer with deep neural networks," Nature , vol. 542, no. 7639, pp. 115–118, 2017, doi: 10.1038/nature21056. S. P. A. Claret, J. P. Dharmian, and A. M. Manokar, "Artificial intelligence-driven enhanced skin cancer diagnosis: leveraging convolutional neural networks with discrete wavelet transformation," Egypt. J. Med. Hum. Genet. , vol. 25, Article 50, 2024, doi: 10.1186/s43042-024-00522-5. F. Mahmud, M. M. Mahfiz, M. Z. I. Kabir, and Y. Abdullah, "An Interpretable Deep Learning Approach for Skin Cancer Categorization," arXiv preprint , arXiv:2312.10696, Dec. 2023. Financial Times , "AI Breakthrough Raises Hopes for Better Cancer Diagnosis," Financial Times , Sep. 4, 2024. C.-E. A. Tai, E. Janes, C. Czarnecki, and A. Wong, "Double-Condensing Attention Condenser: Leveraging Attention in Deep Learning to Detect Skin Cancer from Skin Lesion Images," arXiv preprint , arXiv:2311.11656, Nov. 2023. M. S. Akter, H. Shahriar, S. Sneha, and A. Cuzzocrea, "Multi-class Skin Cancer Classification Architecture Based on Deep Convolutional Neural Network," arXiv preprint , arXiv:2303.07520, Mar. 2023. G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, Densely Connected Convolutional Networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4700–4708. Keras Documentation, “ReduceLROnPlateau,” Keras.io, [Online]. Available: https://keras.io/api/callbacks/reduce_lr_on_plateau/. [Accessed: Apr. 13, 2025]. T. Fawcett, "An introduction to ROC analysis," Pattern Recognition Letters, vol. 27, no. 8, pp. 861–874, 2006, doi: 10.1016/j.patrec.2005.10.010. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted 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-6473686","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":444526986,"identity":"55c47792-bbba-4aae-9ed7-a88c2c7cab32","order_by":0,"name":"Sukh Singh Oberoi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA50lEQVRIie2OsarCMBSGTyl08gEiIr5CRLguoq+SQ6EuPkDHQCHdHEXBh3B0PBLQJegqdFEE54KLgqCBbiLpvdsd8sEJ//B/5AfweP4l4eaEz4ENgYTqrYKDKOZnmVRl+p3S+Gmepa4yQX0fOrkSDNeHUX+ms1sJg/aKQnVyKdxo4mgKXBaoGEHSW1GQc6fCEiFEVAjWQmWHabSKYs5h8wknEe1HVslKgle9AsdJV6KiYN5CaYdRvcLNNgZhYrSKYobHvYWuG5Znu8c9Hdph40uZpsP2dJdf3cM+PrUX/qHv8Xg8nu+8AdWLUyYEBi2/AAAAAElFTkSuQmCC","orcid":"","institution":"SRM Institute of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Sukh","middleName":"Singh","lastName":"Oberoi","suffix":""},{"id":444526988,"identity":"17cdee50-1a64-45ce-9a09-cdecc0757a73","order_by":1,"name":"Ishan Sandhir","email":"","orcid":"","institution":"SRM Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Ishan","middleName":"","lastName":"Sandhir","suffix":""},{"id":444526989,"identity":"57268eab-6c0b-44c2-aec3-2026012c9305","order_by":2,"name":"Rajalakshmi M.","email":"","orcid":"","institution":"SRM Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Rajalakshmi","middleName":"","lastName":"M.","suffix":""}],"badges":[],"createdAt":"2025-04-17 17:08:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6473686/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6473686/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81204599,"identity":"c2613e09-3ab3-4470-925a-af4a565079ce","added_by":"auto","created_at":"2025-04-23 11:58:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":34737,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation Loss over Epochs\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/1c513627d4aa0f809db042bd.png"},{"id":81204600,"identity":"1d29d089-9dab-4fc9-a0dd-d5d4a92b011d","added_by":"auto","created_at":"2025-04-23 11:58:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":36156,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation Accuracy over Epochs\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/e602f8a98e09572d078e38f8.png"},{"id":81204603,"identity":"dbb1711d-381a-41bc-8219-9f27c2534a3a","added_by":"auto","created_at":"2025-04-23 11:58:01","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":63805,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/d0302f4c838f5de1096f668d.png"},{"id":81204606,"identity":"5c8ed269-a7e9-4bf6-8bbc-7f779c61eda3","added_by":"auto","created_at":"2025-04-23 11:58:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":63323,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision, Recall, F1-Score Graph\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/9bf79c2a21bbe7c95a686049.png"},{"id":81205119,"identity":"b0989a1c-63b6-42fa-bb06-93f038000575","added_by":"auto","created_at":"2025-04-23 12:06:01","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":28635,"visible":true,"origin":"","legend":"\u003cp\u003ePer-Class Accuracy of the model\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/f10986430873516afc880627.png"},{"id":81205120,"identity":"158d8eb1-54cf-40cd-a1fe-ff92f169a495","added_by":"auto","created_at":"2025-04-23 12:06:01","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":88835,"visible":true,"origin":"","legend":"\u003cp\u003eROC Curve\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/f45d028c105e64a33d490796.png"},{"id":81204607,"identity":"3e8a1f49-4877-4edc-8242-2487ae91e5f9","added_by":"auto","created_at":"2025-04-23 11:58:01","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":110779,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision-Recall Curve\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/a47d68c39c1783c62843e8f6.png"},{"id":87480818,"identity":"0b19f9e2-6114-4507-9e47-5d24b85241af","added_by":"auto","created_at":"2025-07-24 09:55:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":714786,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6473686/v1/22ce347c-afbd-4db0-a777-08473adcfcf0.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Skin Cancer Classification from Dermatoscopic Images Using Deep Learning Techniques","fulltext":[{"header":"I. INTRDOUCTION","content":"\u003cp\u003eSkin cancer impacts a significant population and can escalate in severity if not identified promptly. Many of these cases are associated with UV radiation exposure, especially in fair skinned people [1]. Melanoma is known to be the most dangerous of all the types. It can spread quickly and cause serious complications or death unless it is found in time [1].\u003c/p\u003e \u003cp\u003eThat\u0026rsquo;s why early detection is so important. However, in many parts of the world, especially rural or low resource areas, people are not always able to access dermatologists or screening tools. Diagnosis in most cases is based on visual inspection with tools such as dermoscopy, which is not always reliable. The experience of the doctor also plays a very big role in it and that can lead to inconsistency [2].\u003c/p\u003e \u003cp\u003eArtificial intelligence has begun to have a more prominent role in healthcare in recent years. In particular, deep learning has strong results in medical imaging. CNNs are very popular because they can learn \u0026ldquo;patterns directly from image data without the need for manual feature selection. Doctors can use these models to make faster\u0026rdquo; and more consistent decisions when they are viewing skin lesions [3].\u003c/p\u003e \u003cp\u003eIn this study, we explore a CNN based approach to classify various types \u0026ldquo;of skin cancer from dermatoscopic images from the ISIC dataset.\u0026rdquo; Preprocessing and augmentation techniques are used to train the model so that it can generalize better. We aim to develop a system that can be used for early diagnosis and is applicable in places where expert resources are scarce.\u003c/p\u003e"},{"header":"II.\tLITERATURE REVIEW","content":"\u003cp\u003eMany researchers have looked into using deep learning and AI for identifying skin cancer. Dildar et al. [1] provided an overview of different neural network models for this purpose. They explained that \u0026ldquo;deep learning models can improve diagnosis if trained with diverse and high-quality datasets\u0026rdquo; [1]. However, they also mentioned issues with real-time use due to \u0026ldquo;high computational demands and limited dataset diversity\u0026rdquo; [1].\u003c/p\u003e \u003cp\u003eMazhar et al. [2] developed a model that combines CNN and SVM techniques to improve classification. They wrote that \u0026ldquo;SVM when combined with CNN showed better generalization performance\u0026rdquo; [2]. But their method had problems detecting rare diseases and produced more false positives.\u003c/p\u003e \u003cp\u003eIn another study, Esteva et al. [3] used popular deep neural networks like VGG16 and ResNet to identify skin cancer. They found that \u0026ldquo;deep neural networks achieved dermatologist-level accuracy\u0026rdquo; [3]. Still, they admitted the system was less reliable when tested with images of people from different skin tones.\u003c/p\u003e \u003cp\u003eClaret et al. [4] improved detection by applying wavelet transformation before feeding data into CNNs. According to them, this helped the model pick up useful features. They mentioned, however, that their approach \u0026ldquo;faced overfitting issues due to small sample sizes\u0026rdquo; [4].\u003c/p\u003e \u003cp\u003eMahmud et al. [5] proposed a model that used attention layers to make the predictions easier to understand. They said the attention mechanism \u0026ldquo;increased transparency of the system\u0026rdquo; [5]. But like others, their model still needed powerful hardware to run effectively.\u003c/p\u003e \u003cp\u003eRecent research has also focused on making skin cancer detection more practical and privacy-friendly. Some teams explored federated learning to avoid sharing patient data [6], while others added attention blocks to better focus on lesions [7]. Akter et al. [8] created a CNN-based model that could classify multiple types of skin cancer and said it \u0026ldquo;achieved competitive performance across different classes\u0026rdquo; [8].\u003c/p\u003e"},{"header":"III. PROPOSED METHODOLOGY","content":"\u003cp\u003eThe classification pipeline consists of several stages: data preprocessing, augmentation, CNN-based model construction, training, and evaluation. A custom convolutional neural network is developed to tell between malignant and benign skin lesions. Each stage is designed to optimize performance while minimizing overfitting, given the class imbalance and limited data variability.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA. Dataset Description\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis study uses data from the International Skin Imaging Collaboration (ISIC) archive, a widely recognized benchmark for skin lesion analysis. It comprises high-resolution dermoscopic images annotated and verified by medical experts. For this work, only two broad categories are considered: benign and malignant.\u003c/p\u003e\n\u003cp\u003eWithin these categories, the dataset includes a variety of skin lesion types:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eBenign: Melanocytic nevi, seborrheic keratoses, and dermatofibromas\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eMalignant: Melanomas, basal cell carcinomas, and actinic keratoses\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThese subtypes introduce substantial visual diversity, making the classification task more challenging and realistic. The dataset also provides extra details like type of lesion, diagnosis, and patient details. However, in this study, only the image data and labels were used.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. Data Preprocessing and Augmentation\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrior to training, all images were standardized to ensure consistency in model input. Each image was resized to 224\u0026times;224 pixels, a common input dimension for convolutional neural networks, and pixel intensities were normalized between 0 and 1 to improve numerical stability during training.\u003c/p\u003e\n\u003cp\u003eTo address class imbalance\u0026mdash;where benign lesions significantly outnumber malignant ones\u0026mdash;the dataset was balanced by undersampling the benign class. This ensures that the model gets a balanced set of samples from both classes, reducing bias in predictions and improving sensitivity to malignant cases.\u003c/p\u003e\n\u003cp\u003eTo improve generalization and reduce overfitting, data augmentation was employed on the training set. Augmentation techniques included:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eRandom rotations\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eHorizontal and vertical flipping\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eZooming\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eWidth and height shifts\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThese transformations were applied in real-time during training using data generators. By exposing the model to varied representations of the same image, augmentation enhances its ability to learn invariant and robust features, which is critical in medical imaging tasks with limited datasets.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eC. Model Architecture\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe proposed model leverages transfer learning by utilizing DenseNet201, \u0026ldquo;a convolutional neural network pre-trained on the ImageNet dataset, to extract features\u0026rdquo; [9, p. 4700]. As Huang et al. describe, \u0026ldquo;each layer obtains additional inputs from all preceding layers and passes on its own feature-maps to all subsequent layers\u0026rdquo; [9, p. 4700]. This dense connectivity \u0026ldquo;encourages feature reuse and substantially reduces the number of parameters\u0026rdquo; [9, p. 4700] compared to traditional CNNs.\u003c/p\u003e\n\u003cp\u003eThe model uses a DenseNet201 backbone pre-trained on ImageNet as a feature extractor, followed by a small custom classifier consisting of a Dropout layer, a Dense layer with 512 units, and a Softmax output layer with 9 classes. This structure balances computational efficiency with strong performance on the skin cancer classification task. A full architecture summary is provided in Table I.\u003c/p\u003e\n\u003cp\u003eTable 1. Model Architecture Summary\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable id=\"Taba\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eLayer Type\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eOutput Shape\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eParameters\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eActivation\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDenseNet201 (pretrained)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(2, 3, 1920)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18,321,984\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eReLU\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFlatten\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(11520)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDropout (0.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(11520)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDense (512 units)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(512)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5,898,752\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eReLU\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDense (9 units)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4,617\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSoftmax\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003eTotal params: 24,225,353 (92.41 MB)\u003c/p\u003e\n \u003cp\u003eTrainable params: 23,996,297 (91.54 MB)\u003c/p\u003e\n \u003cp\u003eNon-trainable params: 229,056 (894.75 KB)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cem\u003eD. Training Configuration\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u0026ldquo;The model was compiled using the Stochastic Gradient Descent (SGD) optimizer with a learning rate of 0.001 and a momentum of 0.9, which was found to provide stable convergence during experimentation. The categorical cross-entropy loss function was used due to the multi-class nature of the classification task.\u003c/p\u003e\n\u003cp\u003eTraining was done for 50 epochs with a batch size of 32, and the data was split into 80% for training and 20% for validation.\u003c/p\u003e\n\u003cp\u003eTo support model convergence, a learning rate reduction strategy was also applied by monitoring validation accuracy, reducing the learning rate by a factor of 0.5 with a patience of 3 epochs, and setting a minimum learning rate threshold of 0.00001.\u003c/p\u003e\n\u003cp\u003eTraining was conducted using a single NVIDIA T4 GPU, and the model required approximately 45 minutes, depending on system load.\u003c/p\u003e\n\u003cp\u003eAs shown in Figures 1 and 2, the model demonstrates a steady increase in training and validation accuracy along with a consistent decrease in loss, indicating effective training convergence.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eE. Evaluation Metrics\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eWe evaluated the model using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix metrics, as detailed in Section IV.\u003c/p\u003e"},{"header":"IV. RESULT AND DISCUSSION","content":"\u003cp\u003eA. \u003cem\u003eResult and Evaluation of Model\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis section presents the experimental results obtained from training and evaluating the proposed model on the skin cancer classification task. Multiple performance metrics are analyzed to assess overall accuracy, class-wise behavior, and model reliability, especially under the medical constraints of misdiagnosis and class imbalance.\u003c/p\u003e\n\u003cp\u003e1) \u003cem\u003eAccuracy\u003c/em\u003e: Used during training and validation to monitor overall performance. \u0026ldquo;The model achieved a training accuracy of 99.41%, a validation accuracy of 82.40%, and a final test accuracy of 84.00%.\u0026rdquo; Accuracy curves across epochs are shown in Fig.\u0026nbsp;2.\u003c/p\u003e\n\u003cp\u003e2) \u003cem\u003eConfusion Matrix\u003c/em\u003e: A 9\u0026times;9 \u0026ldquo;confusion matrix was generated to\u0026rdquo; visualize class-wise prediction performance. This helped identify misclassifications between similar skin cancer types. (Fig. 3)\u003c/p\u003e\n\u003cp\u003e3) \u003cem\u003ePrecision, Recall, F1-Score\u003c/em\u003e: A classification report was computed for all nine classes. These metrics are crucial for evaluating performance in tasks with imbalanced classes, particularly in medical diagnostics. As described by Fawcett, \u0026quot;Precision measures the proportion of positive predictions that are actually correct, recall indicates the proportion of actual positives that were correctly identified, and the F1-score is the harmonic mean of precision and recall\u0026quot; [11, p. 862].\u003c/p\u003e\n\u003cp\u003e4) \u003cem\u003ePer-Class Accuracy\u003c/em\u003e: Individual class-wise accuracy values were also computed to analyze how well the model performs across specific skin cancer types. This is crucial in medical applications where certain types may be harder to detect. (Fig. 5)\u003c/p\u003e\n\u003cp\u003e5) \u003cem\u003e\u0026ldquo;ROC (Receiver Operating Characteristic) Curve\u003c/em\u003e: ROC\u0026rdquo; curves were plotted for each class by computing one-vs-rest scores (Fig. 6). As explained in the literature, \u0026ldquo;the area under the ROC curve (AUC) provides an aggregate measure of performance across all possible classification thresholds\u0026rdquo; [11].\u003c/p\u003e\n\u003cp\u003e6) \u003cem\u003ePrecision\u0026ndash;Recall (PR) Curve\u003c/em\u003e: PR curves were plotted to evaluate performance under class imbalance. These are especially useful when false positives carry lower risk than false negatives, as is often the case in medical screening tasks (Fig. 7).\u003c/p\u003e\n\u003cp\u003e7) \u003cem\u003eLearning Rate Scheduling\u003c/em\u003e: Keras\u0026rsquo; ReduceLROnPlateau callback was employed which \u0026ldquo;Reduce learning rate when a metric has stopped improving\u0026rdquo; [10] as \u0026ldquo;models often benefit from reducing the learning rate by a factor of 2\u0026ndash;10 once learning stagnates\u0026rdquo; [10].\u003c/p\u003e\n\u003cp\u003eB. \u003cem\u003eDiscussion\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u0026ldquo;The model demonstrates strong generalization with a high training accuracy of 99.41% and a reasonably good test accuracy of 84.00%. However, the performance gap between training and validation suggests potential overfitting, which could be mitigated further with more data or stronger regularization. Certain classes, such as melanoma and\u0026rdquo; basal cell carcinoma, showed lower precision and recall, highlighting the challenge of distinguishing between visually similar lesion types. Future work could explore ensemble models, better augmentation, or transfer learning from larger dermatology datasets to improve robustness.\u003c/p\u003e"},{"header":"V. CONCLUSION","content":"\u003cp\u003eThis study presents a deep learning-based approach to classify multiple types of skin cancer using dermatoscopic images. The model was trained and tested on the ISIC dataset, which includes nine skin lesion categories. Through extensive preprocessing, data augmentation, and a well-designed convolutional neural network, the model reached a final test accuracy of 84.00%, with promising class-wise performance across most categories.\u003c/p\u003e\n\u003cp\u003eEvaluation using confusion matrix, ROC and precision–recall curves, and class-specific metrics such as F1-score confirmed the model's effectiveness, while also highlighting the challenges in distinguishing visually similar lesion types. The results show that deep learning has strong potential in aiding early detection of skin cancer, particularly as a diagnostic support tool for dermatologists.”\u003c/p\u003e"},{"header":"VI. FUTURE WORK","content":"\u003cp\u003eAlthough the proposed model shows promising results, several enhancements can be made to improve its performance, accuracy, reliability and real-world applicability:\u003c/p\u003e \u003cp\u003eA. \u003cem\u003eLarger and More Diverse Datasets\u003c/em\u003e: Incorporating more varied and real-world clinical images, including higher-resolution samples and other demographics, can improve generalization.\u003c/p\u003e \u003cp\u003eB. \u003cem\u003eTransfer Learning and Model Ensembling\u003c/em\u003e: Using pretrained medical imaging models or combining multiple architectures may enhance feature extraction and robustness.\u003c/p\u003e \u003cp\u003eC. \u003cem\u003eExplainable AI (XAI)\u003c/em\u003e: Integrating model interpretability techniques (e.g., Grad-CAM or SHAP) can help clinicians trust the model's predictions and understand decision boundaries.\u003c/p\u003e \u003cp\u003eD. \u003cem\u003eMobile/Edge Deployment\u003c/em\u003e: Optimizing the model for low-latency edge devices could make it suitable for use in telemedicine and rural healthcare settings.\u003c/p\u003e \u003cp\u003eE. \u003cem\u003eClinical Validation\u003c/em\u003e: Collaborating with dermatologists to validate the model in real-world clinical workflows would be a critical next step toward deployment.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eS.O. and I.S. wrote the main manuscript text and S.O. added and reviewed citations and references. I.S. prepared introduction. R.M. guided us throughout the project and reviewed the manuscript.\u003c/p\u003e\u003ch2\u003e \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eAcknowledgments\u003c/span\u003e \u003c/h2\u003e \u003cp\u003e\u0026ldquo;We sincerely thank the faculty of the Department of Computing Technologies at SRM Institute of Science and Technology for their continuous support and insightful feedback during the course of this project. We are especially grateful to Dr. Rajalakshmi M, Assistant Professor, for her valuable guidance and encouragement throughout the research.\u0026rdquo;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eM. Dildar, S. Akram, M. Irfan, et al., \u0026quot;Skin Cancer Detection: A Review Using Deep Learning Techniques,\u0026quot; Int. J. Environ. Res. Public Health, vol. 18, no. 10, p. 5479, May 2021, doi: 10.3390/ijerph18105479.\u003c/li\u003e\n \u003cli\u003eT. Mazhar, I. Haq, A. Ditta, \u003cem\u003eet al\u003c/em\u003e., \u0026quot;The Role of Machine Learning and Deep Learning Approaches for the Detection of Skin Cancer,\u0026quot; \u003cem\u003eHealthcare\u003c/em\u003e, vol. 11, no. 3, p. 415, 2023, doi: 10.3390/healthcare11030415.\u003c/li\u003e\n \u003cli\u003eA. Esteva, B. Kuprel, R. A. Novoa, \u003cem\u003eet al\u003c/em\u003e., \u0026quot;Dermatologist-level classification of skin cancer with deep neural networks,\u0026quot; \u003cem\u003eNature\u003c/em\u003e, vol. 542, no. 7639, pp. 115\u0026ndash;118, 2017, doi: 10.1038/nature21056.\u003c/li\u003e\n \u003cli\u003eS. P. A. Claret, J. P. Dharmian, and A. M. Manokar, \u0026quot;Artificial intelligence-driven enhanced skin cancer diagnosis: leveraging convolutional neural networks with discrete wavelet transformation,\u0026quot; \u003cem\u003eEgypt. J. Med. Hum. Genet.\u003c/em\u003e, vol. 25, Article 50, 2024, doi: 10.1186/s43042-024-00522-5.\u003c/li\u003e\n \u003cli\u003eF. Mahmud, M. M. Mahfiz, M. Z. I. Kabir, and Y. Abdullah, \u0026quot;An Interpretable Deep Learning Approach for Skin Cancer Categorization,\u0026quot; \u003cem\u003earXiv preprint\u003c/em\u003e, arXiv:2312.10696, Dec. 2023.\u003c/li\u003e\n \u003cli\u003e\u003cem\u003eFinancial Times\u003c/em\u003e, \u0026quot;AI Breakthrough Raises Hopes for Better Cancer Diagnosis,\u0026quot; \u003cem\u003eFinancial Times\u003c/em\u003e, Sep. 4, 2024.\u003c/li\u003e\n \u003cli\u003eC.-E. A. Tai, E. Janes, C. Czarnecki, and A. Wong, \u0026quot;Double-Condensing Attention Condenser: Leveraging Attention in Deep Learning to Detect Skin Cancer from Skin Lesion Images,\u0026quot; \u003cem\u003earXiv preprint\u003c/em\u003e, arXiv:2311.11656, Nov. 2023.\u003c/li\u003e\n \u003cli\u003eM. S. Akter, H. Shahriar, S. Sneha, and A. Cuzzocrea, \u0026quot;Multi-class Skin Cancer Classification Architecture Based on Deep Convolutional Neural Network,\u0026quot; \u003cem\u003earXiv preprint\u003c/em\u003e, arXiv:2303.07520, Mar. 2023.\u003c/li\u003e\n \u003cli\u003eG. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, Densely Connected Convolutional Networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4700\u0026ndash;4708.\u003c/li\u003e\n \u003cli\u003eKeras Documentation, \u0026ldquo;ReduceLROnPlateau,\u0026rdquo; Keras.io, [Online]. Available: https://keras.io/api/callbacks/reduce_lr_on_plateau/. [Accessed: Apr. 13, 2025].\u003c/li\u003e\n \u003cli\u003eT. Fawcett, \u0026quot;An introduction to ROC analysis,\u0026quot; Pattern Recognition Letters, vol. 27, no. 8, pp. 861\u0026ndash;874, 2006, doi: 10.1016/j.patrec.2005.10.010.\u003cstrong\u003e\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Skin Disease Detection, Deep Learning, Dermoscopic Images, AI in Healthcare, CNN Classification ","lastPublishedDoi":"10.21203/rs.3.rs-6473686/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6473686/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSkin cancer affects a large number of people and can become serious if not detected early. In many places, timely diagnosis is challenging due to a lack of specialists. Also, skin lesions often look very similar, which makes visual diagnosis tricky. This study looks at using deep learning to help with that. We trained a convolutional neural network model to classify several types of skin cancer using images from the ISIC dataset. Before training, the images were pre-processed and augmented to improve model performance. The model was tested using common metrics like accuracy, precision, recall, and ROC-AUC. The results indicate that AI can support doctors by making skin cancer detection faster and more reliable.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"Skin Cancer Classification from Dermatoscopic Images Using Deep Learning Techniques","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-23 11:57:56","doi":"10.21203/rs.3.rs-6473686/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e214a4c7-f3e7-4bd3-9d01-d9c05ceec483","owner":[],"postedDate":"April 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-07-24T09:38:55+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-23 11:57:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6473686","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6473686","identity":"rs-6473686","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}