FAT-Net: Feature adaptive transformers for automated skin lesion segmentation

Area of Science:

• Medical image analysis
• Artificial intelligence in dermatology
• Computational pathology

Background:

• Accurate skin lesion segmentation is crucial for melanoma diagnosis and quantitative analysis.
• Challenges include scale variations, irregular shapes, and blurred boundaries in dermoscopic images.
• Traditional Convolutional Neural Networks (CNNs) struggle with capturing global context, limiting segmentation performance.

Purpose of the Study:

• To develop an advanced deep learning model for improved skin lesion segmentation.
• To address the limitations of existing CNN-based methods in capturing long-range dependencies.
• To enhance feature fusion and reduce background noise for more precise segmentation.

Main Methods:

• Proposed a novel Feature Adaptive Transformer Network (FAT-Net) using an encoder-decoder architecture.
• Integrated a transformer branch to capture global context and long-range dependencies.
• Employed a memory-efficient decoder and feature adaptation module for enhanced feature fusion.

Main Results:

• FAT-Net demonstrated superior performance on four public skin lesion segmentation datasets (ISIC 2016, 2017, 2018, PH2).
• Ablation studies confirmed the effectiveness of the feature adaptive transformers and memory-efficient strategies.
• The model achieved state-of-the-art accuracy and inference speed compared to existing methods.

Conclusions:

• FAT-Net offers a significant advancement in skin lesion segmentation accuracy and efficiency.
• The proposed architecture effectively handles challenges like scale variation and blurred boundaries.
• This method holds promise for improving automated melanoma analysis and diagnosis.