MLFFM-SegDiff: New Diffusion Model Boosts Skin Lesion Segmentation Accuracy

Skin lesion segmentation is critical for computer-aided dermatology, but dermoscopic images suffer from blurred boundaries, low contrast, shape variations, and artifacts like hair and shadows. Diffusion models have shown promise in medical image segmentation due to their progressive denoising and distribution modeling. However, existing diffusion-based methods struggle with cross-level feature interaction and boundary detail recovery, limiting their accuracy on challenging cases.

To overcome these limitations, the team proposes MLFFM-SegDiff, built on a diffusion framework with three key innovations: a dual-path U-Net encoder that enhances interaction between noisy mask features and dermoscopic image features; a Multi-Level Feature Fusion Module (MLFFM) that improves skip connections via attention, scale alignment, and adaptive cross-level fusion; and a boundary-sensitive loss function. These designs let the decoder jointly leverage shallow boundary cues and deep semantic representations, improving mask reconstruction quality. On the ISIC2018, PH2, and HAM10000 datasets, MLFFM-SegDiff achieves an average Jaccard index of 0.8546 and Dice coefficient of 0.9207, outperforming DermoSegDiff, U-Net, and SwinUNETR across Accuracy, F1-score, Jaccard, Recall, and Dice. The code will be released after publication.