ScaleFusionNet: transformer-guided multi-scale feature fusion for skin lesion segmentation.

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of skin lesions is vital for melanoma diagnosis.
• Challenges include blurred boundaries, color variations, and irregular shapes.

Purpose of the Study:

• To develop an advanced deep learning model for precise skin lesion segmentation.
• To improve the accuracy and efficiency of quantitative analysis in dermatological imaging.

Main Methods:

• Proposed ScaleFusionNet, a hybrid model integrating Cross-Attention Transformer Module (CATM) and adaptive fusion block (AFB).
• CATM uses Swin transformer blocks and Cross Attention Fusion (CAF) for refined feature fusion.
• AFB employs Swin Transformer attention and deformable convolution for adaptive feature extraction.

Main Results:

• Achieved high Dice scores: 92.94% (ISIC-2016), 91.80% (ISIC-2018), and 95.37% (HAM10000).
• Demonstrated significant performance improvements over state-of-the-art methods on the PH² dataset.
• Successfully refined lesion boundaries and preserved fine-grained details.

Conclusions:

• ScaleFusionNet effectively enhances feature extraction and fusion for improved skin lesion segmentation.
• The model shows significant potential for clinical applications in melanoma detection and analysis.
• Publicly available code facilitates further research and development.