Diagnosis-Stratified U-Net for Dermoscopic Lesion Segmentation: A Reproducible Baseline with Boundary-Aware Evaluation

Background: Automated skin lesion segmentation is essential to computer-aided diagnosis, with existing approaches commonly neglecting boundary accuracy and performance variability based on morphology, which constrains clinical applications to surgical planning. Methods: We introduce a diagnosis-conscious U-Net model tested on the ISIC 2018 dataset (519 test images) with stratified partitioning to avoid diagnostic leakage. The pipeline uses error-graded augmentation and end-to-end assessment based on overlap measures (Dice, IoU), boundary-sensitive Hausdorff distance, and per-diagnosis assessment between melanoma, nevus, and seborrheic keratosis. A systematic comparison is made between baseline configurations and stratified splitting, full augmentation and Dice loss optimization in an ablation study. Results: The model attained a Dice coefficient of 0.871 ± 0.138, sensitivity of 0.921 and specificity of 0.968. The mean Hausdorff distance was 24.65 ± 28.94 pixels (~1.23 mm) which showed varying boundary accuracy. The heterogeneity of performance was demonstrated with per-diagnosis analysis: nevus (0.890), melanoma (0.840), and seborrheic keratosis (0.779 ± 0.230). The stratified splitting enhanced the minority-class F1 by 0.010, whereas aggressive augmentation harmed the performance, implying that the optimal performance requires dataset-specific optimization. Conclusion: In morphology-sensitive failure modes that are masked by aggregate measures, diagnosis-stratified evaluation can be used. Although it is appropriate to screen, the 1.23 mm deviation in a boundary requires adjustment to margincritical surgical uses. To facilitate reproducible benchmarking and improvements in the future with morphologybased adjustments, we publish code and pretrained weights.