Apppublic

usamyy/isic-skin-lesion-segmentation

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

App README

ISIC Skin Lesion Segmentation — SwinUNet

A SwinUNet model trained from scratch on a 640-image subset of the ISIC 2018 Lesion Boundary Segmentation dataset. Submitted as part of an MS Data Science assignment in Medical Image Computing.

Model

—Architecture: SwinUNet (Cao et al., 2022)
—Parameters: ~27M
—Training: from scratch, 50 epochs, Dice + BCE loss, AdamW + cosine LR
—Test Dice: 0.8731 | IoU: 0.7967 | Precision: 0.8947 | Recall: 0.8960

Research demo only. NOT for clinical use.