HCViT-Net: Hybrid CNN and multi scale query transformer network for dermatological image segmentation

Area of Science:

• Artificial Intelligence
• Medical Image Analysis
• Computer Vision

Background:

• Dermoscopic lesion segmentation is vital in dermatology.
• Current methods face challenges in integrating global context and local details efficiently for clinical use.

Purpose of the Study:

• To develop a lightweight model for dermoscopic lesion segmentation.
• The model aims to capture long-range spatial dependencies and fine-grained boundary details.
• Achieve a favorable accuracy-efficiency trade-off for practical clinical deployment.

Main Methods:

• Proposed a lightweight hybrid model, HCViT-Net, with an encoder-decoder architecture.
• Incorporated a multi-scale query transformer (MSQFormer) for efficient global context capture.
• Introduced a wavelet-guided attention refinement module (WARM) to enhance boundary details.

Main Results:

• Achieved high mean intersection-over-union (mIoU) scores of 87.76% (ISIC 2017) and 87.45% (ISIC 2018).
• The model has only 5.76M parameters and 7.51 GFLOPs, indicating low computational cost.
• Demonstrated competitive performance compared to existing methods with significantly reduced computational requirements.

Conclusions:

• HCViT-Net offers an excellent accuracy-efficiency trade-off for dermoscopic lesion segmentation.
• The model improves segmentation accuracy with a low computational footprint.
• Shows strong potential for integration into clinical dermatology workflows.