بخش‌بندي ضايعات پوستي با استفاده از شبكه؜‌هاي عصبي عميق

Skin cancer is among the most common an‎d deadly cancers worldwide. Accurate segmentation of skin lesions plays a crucial role in computer-aided skin cancer diagnosis systems; however, challenges such as visual variability an‎d irregular lesion boundaries limit the performance of existing models. This study aims to improve segmentation accuracy while maintaining computational complexity at a level suitable for real-time applications. To this end, two deep learning-based models are proposed. The first model, inspired by UCM-NetV2, employs a dual-path encoder based on MobileViT-S an‎d a semantics an‎d detail infusion module. By leveraging test-time augmentation an‎d a small-lesion detection an‎d cropping mechanism, this model demonstrates superior capability in segmenting challenging samples compared to the baseline model. It achieved DSC of 92.66% an‎d 91.04% on the ISIC2017 an‎d ISIC2018 datasets, respectively, outperforming the baseline model while maintaining low computational complexity. The second model is an enhanced version of the novel DFF-UNet, in which the dual-path encoder, feature fusion module, improved decoder, an‎d bridge module have been redesigned an‎d optimized. The results indicate that the second proposed model achieved DSC improvements of 4.17% an‎d 2.35% on ISIC2017 an‎d ISIC2018 datasets, respectively, compared to the baseline. Comparative performance analysis shows that both proposed models outperform their respective baseline models, with the first model being lighter, while the second model achieves the highest accuracy in most scenarios. Overall, the findings of this study confirm the effectiveness of the proposed architectures in improving skin lesion segmentation.

نادر كريمي

شادرخ سماوي

حميدرضا حكيم داودي , مينا اميري