HyperFusionNet combines vision transformer for early melanoma detection and precise lesion segmentation

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Early and accurate melanoma diagnosis is crucial but challenging due to lesion heterogeneity and diagnostic tool limitations.
• Current methods often struggle with simultaneous classification and segmentation of skin lesions.

Purpose of the Study:

• To introduce HyperFusion-Net, a novel hybrid deep learning architecture for simultaneous melanoma classification and lesion segmentation.
• To evaluate the performance of HyperFusion-Net against state-of-the-art models on public dermoscopic image datasets.

Main Methods:

• Developed HyperFusion-Net, integrating a Multi-Path Vision Transformer (MPViT) with an attention U-Net and a mutual attention fusion block.
• Trained and evaluated the model on over 60,000 dermoscopic images from four public ISIC datasets.
• Applied preprocessing techniques including hair removal, clipping, and normalization for enhanced robustness.

Main Results:

• HyperFusion-Net achieved superior classification accuracy (93.24%) and AUC (95.80%).
• The model demonstrated excellent segmentation performance with a Dice coefficient of 0.945 (ISIC 2024).
• Outperformed existing models like U-Net, DeepLabV3+, TransUNet, and Swin-UNet in both tasks.

Conclusions:

• HyperFusion-Net effectively integrates transformer and U-Net features for accurate melanoma diagnosis.
• The model shows strong generalizability across diverse datasets and imaging conditions, offering computational efficiency.
• This hybrid approach represents a significant advancement in automated skin lesion analysis for improved melanoma detection.