A Multitask Deep Learning Approach for Tumor Segmentation and Morphological Quantification in Acral Melanoma...

This study presents an integrated multitask deep learning framework for the automated analysis of acral melanoma from whole-slide images (WSIs). We constructed a multiscale WSI dataset with pixel-level annotations and developed a pipeline that sequentially performs classification, segmentation, and morphological quantification. An optimized ResNet50 classifier first screens for malignancy. For malignant cases, our novel D3C-RS-Unet, which integrates dual dilated dynamic convolution and channel attention mechanisms, performs precise tumor segmentation. Finally, key morphological indicators are automatically extracted. Using patient-level fivefold cross-validation, the classifier achieved an accuracy of 0.8947 ± 0.0098. The D3C-RS-Unet significantly outperformed baseline and state-of-the-art models (Dice: 0.8427 ± 0.0057; IoU: 0.8053 ± 0.0071; p_adj < 0.05). Extracted morphological metrics showed excellent agreement with manual annotations (ICCs > 0.93). This end-to-end framework provides objective, reproducible quantitative data, demonstrating significant potential to assist pathological diagnosis and standardize reporting in acral melanoma.