Processing an‎d Detection of Skin Cancer Based on Neural Networks

10/22/2025 12:00:00 AM

Skin cancer is the most common type of cancer worldwide an‎d is preventable. The use of deep learning networks for early detection of these lesions can play a key role in accelerating the treatment process an‎d increasing patient survival rates. In this study, a comprehensive an‎d multimodal framework was designed an‎d eva‎luated for the classification of skin lesions, with the aim of improving accuracy an‎d clinical safety. The main challenges in this field are the high visual similarity between benign an‎d malignant lesions, as well as the data imbalance in stan‎dard datasets. To address these challenges, a complementary dual-branch architecture was proposed: an accuracy-oriented branch to maximize overall accuracy, an‎d a balance-oriented branch to improve the recognition of minority classes. This architecture was further enhanced with an attention module to enable the intelligent integration of clinical metadata.‌‌This framework was tested in two main phases. In the first phase, which focused on the HAM10000 dataset, various strategies for han‎dling data imbalance were compared. In this phase, the baseline model achieved an overall accuracy of 93.8%. However, it was shown that by using weighted sampling, the sensitivity for the melanoma class could be improved from 70.5% to 72.3%.‌In the second phase, the model was eva‎luated on the more complex ISIC 2019 dataset using the enhanced architecture. By employing class-weighting in the loss function an‎d the advanced Snapshot ensemble technique, the final model was able to achieve an overall accuracy of 92.4% while increasing the sensitivity (recall) for the melanoma class to 86%‌. The final results show that the proposed approach, particularly in the second phase,‌‌has successfully achieved the optimal balance between statistical accuracy an‎d clinical safety. Achieving 86% sensitivity for melanoma on the challenging ISIC 2019 dataset demonstrates the high effectiveness of this framework as a reliable an‎d responsible diagnostic support tool.

طبقه بندي سرطان , يادگيري عميق

skin cancer , deep learning

Bita Movahedi

Ahmad ayatollahi