A Deep Learning-Based CNN Framework for Automated Skin Disease Classification and Diagnosis

Abstract

Skin disorders are very common worldwide and must be identified quickly and accurately in order to prevent sequelae. A typical diagnosis relies on subjective visual examina- tion by dermatologists, which can vary and difficult in resource limited settings. This contribution overcomes the previously mentioned limitations and uses a deep learning architecture based on convolutional neural networks for diagnosis. (CNNs), for skin lesion detection and recognition. Exploiting high level feature extraction ability, the model is trained on well-known datasetsHAM10000andISIC.Resultsintermsofsalientmetrics, including accuracy, precision, recall, and F1 score establish that the proposed CNN model achieves extensive improvement over traditionalapproachesintermsofdiagnosticspeedandtrustwor- thiness.rate and distills intelligence observed from the pattern in today’sstate-of-the-artartificialintelligence(AI)systems.Forex- ample:Skindiseasesaresufferingavastpercentageoftheworld’s population, where an early and accurate diagnosis is essential to avoid severe after effects. Standard diagnostic procedures that require dermatologist’s examination may subjective, not time effective, and less trust worthy in rural areas. This proposed study presents an automated system based on deep learning to classifyskinlesionsusingCNNarchitecturesfromvisualfeatures. Trained using popular datasets such as HAM10000 and ISIC,thesystem’sperformanceintermsofAccuracy,Precision,Recall and F1-score outperforms classical classifier and this work has potential to help speed up and improve the accuracy of clinical decision-making.

Introduction

Skin diseases are among the most prevalent health conditions worldwide, making early and accurate diagnosis essential, particularly for life-threatening diseases such as melanoma. Traditional diagnostic methods, including visual examination, dermoscopy, and biopsy, are effective but are often time-consuming, subjective, and dependent on specialist availability. Earlier computer-aided diagnostic systems based on handcrafted image features had limited accuracy and poor generalization across different skin types and imaging conditions. Recent advances in deep learning, especially Convolutional Neural Networks (CNNs), have significantly improved skin lesion classification by automatically learning complex features from medical images, enabling dermatologist-level diagnostic performance and supporting accessible, scalable healthcare solutions.

The proposed system introduces a CNN-based automated skin disease detection framework that classifies skin lesions into multiple disease categories using benchmark datasets such as HAM10000, ISIC, and PAD-UFES-20. The methodology includes image validation, resizing, color standardization, normalization, illumination enhancement, artifact removal, denoising, and lesion segmentation to improve image quality. Hierarchical feature extraction through CNN layers captures both low-level textures and high-level lesion characteristics, while fully connected dense layers and a Softmax classifier generate probability-based disease predictions. The model is optimized using techniques such as dropout, batch normalization, and categorical cross-entropy loss to improve robustness and reduce overfitting.

Experimental results demonstrate that the proposed CNN + Dense Model outperforms traditional machine learning methods and existing CNN approaches. It achieves 91.75% accuracy, 90.9% precision, and 90.5% recall, compared to lower performance from Support Vector Machine (82.15% accuracy), Random Forest (84.6%), and conventional CNN models (88.4%). The model exhibits smooth convergence, minimal overfitting, and improved reliability in detecting various skin diseases. Although challenges such as dataset imbalance and computational complexity remain, the proposed framework provides a highly accurate, scalable, and clinically valuable solution for automated skin disease diagnosis, with strong potential for teledermatology and rural healthcare applications.

Conclusion

The current study successfully developed and evaluated a deep learning–based framework using a CNN architecture combinedwithdenselayersforskindiseaseclassification. It also demonstrated how advanced preprocessing, convolu- tional feature extraction, and optimized integration of dense layerssignificantlyenhancediagnosticaccuracy.Theproposed modelachievedanoverallaccuracyof91.75%,withprecision, recall, and F1-scores greater than 90%, on the HAM10000 dataset, thereby confirming its reliability for clinical decision support and its superiority against conventional machine- learning techniques and CNN architectures. The importanceof normalization, augmentation, and class balancing in reduc- ing bias and improving generalization within different lesion categories was also emphasized. While the performance ofthe system was strong, further improvements could be made by expanding the dataset with diverse sources such as theISICorDerm7ptdatasets,consideringtransferlearningmodels like EfficientNet or Vision Transformers to enhance feature extraction, and integrating explainable AI tools like Grad- CAM for enhanced interpretability to dermatologists. Further, real-timedeploymentthroughCADsystemsonmobileorweb- basedsystems,optimizationusinglightweightframeworkslike TensorFlowLite,andintegrationofpatientmetadataformulti- modal diagnosis emerge as promising directions to increase usabilityanddiagnosticdepth.

References

[1] Brinker, F., Hekler, A., Enk, A. H., et al. (2019). Deep learningperformance compared with dermatologists in dermoscopic melanomaclassification. European Journal of Cancer, 119, 57–65. https://doi.org/10.1016/j.ejca.2019.04.001 [2] Brinker, F., Hekler, A., Utikal, J. S., et al. (2019). A CNN trained ondermoscopy performs comparably to dermatologists on clinical images.European Journal of Cancer, 120, 114–120. https://doi.org/10.1016/j.ejca.2019.02.005 [3] Combalia, A., Puig, S., Malvehy, J., et al. (2022). Machine learningvalidation for dermoscopic skin cancer classification: Insights from theISIC Challenge. The Lancet Digital Health, 4(6), e469–e479. https://doi.org/10.1016/S2589-7500(22)00021-8 [4] Esteva, A., Kuprel, B., Novoa, R. A., et al. (2017). Dermatologist-levelskincancerclassificationusingdeepneuralnetworks.Nature,542,115–118. https://doi.org/10.1038/nature21056 [5] Molina-Molina,J.,Lo´pez,J.,&Pe´rez,A.(2020).AhybridCADsystemintegrating fractal analysis and DenseNet-201 for dermoscopy-baseddiagnosis. Applied Sciences, 10(22), 8083–8098. https://doi.org/10.3390/app10175954 [6] Alam, M., Islam, M., & Rahman, M. (2022). A deep learning classifierfor multi-class skin lesion recognition using the HAM10000 dataset.Computers in Biology and Medicine, 145, 105454. https://doi.org/10.3390/diagnostics12092115 [7] Kaur, P., Singh, S., & Sharma, A. (2022). Automated melanoma andbenign lesion classification using DCNNs and dermoscopic images.Sensors, 22(14), 5237. https://doi.org/10.3390/s22031134 [8] Adepu, K., Reddy, P., & Singh, D. (2023). An optimized CNN pipelinefor dermoscopic image classification: SIIM-ISIC 2020 challenge analy-sis. Computers in Biology and Medicine, 153, 106455. https://doi.org/10.1016/j.compbiomed.2023.106571 [9] Venugopal, P., Rao, S., & Kumar, N. (2023). Modified EfficientNetdesign for dermoscopic skin lesion classification. Biomedical SignalProcessing and Control, 86, 104859. https://doi.org/10.1016/j.dajour.2023.100278 [10] Akram, M., Khan, S., & Hussain, M. (2023). A hybrid segmentation–classificationpipelineformelanomadetectionindermoscopy.Biocyber-netics and Biomedical Engineering, 43(2), 534–549. https://doi.org/10.1111/srt.13524 [11] Alshahrani, A., et al. (2024). Hybrid multi-CNN feature fusion fordermoscopic image classification using ISIC-2019 dataset. Healthcare,12(1), 55. https://doi.org/10.1371/journal.pone.0298305 [12] Musthafa, M., et al. (2024). Optimized CNN stacking for efficient andaccurate skin lesion classification. Diagnostics, 14(2), 321. https://doi.org/10.1186/s12880-024-01356-8 [13] Mateen,M.,etal.(2024).AhybriddeeplearningframeworkusingU-Net and Inception–ResNet-v2 for skin lesion segmentation and clas-sification. Diagnostics, 14(3), 417. https://doi.org/10.3390/diagnostics14192242 [14] Al-Waisy, H., et al. (2025). Skin-DeepNet: An efficient framework forearly diagnosis of skin cancer from dermoscopic images. ScientificReports, 15(10), 12562. https://doi.org/10.1038/s41598-025-15655-9 [15] Nawaz, S., et al. (2025). FCDS-CNN: A class-imbalance–aware con-volutional neural network for melanoma detection. Scientific Reports,15(8), 10225. https://doi.org/10.1038/s41598-025-91446-6

Copyright

Copyright © 2026 Ms. J Sirisha, Ilapuram Pavan, Pogaku Jaswanth, Mandrekar Jithender. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.