Polyp Segmentation Using a Hybrid Vision Transformer and a Hybrid Loss Function.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) often develops from adenomatous polyps, making early detection critical for reducing mortality.
• Current colonoscopy screening faces limitations including variable image quality, physician fatigue, and potential for missed diagnoses.
• Computer-aided detection methods exist but often have limitations that hinder widespread clinical adoption.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture for accurate polyp segmentation in colonoscopy images.
• To leverage both high-level semantic and low-level spatial features for improved segmentation performance.
• To introduce a hybrid loss function that enhances regional consistency and reduces segmentation errors.

Main Methods:

• A new segmentation architecture based on residual transformer layers was designed.
• The model integrates high-level semantic and low-level spatial features for comprehensive analysis.
• A novel hybrid loss function combining focal Tversky loss, binary cross-entropy, and Jaccard index was implemented.

Main Results:

• The proposed approach achieved high performance metrics: Dice similarity (0.9048), recall (0.9041), precision (0.9057), and F2 score (0.8993).
• Experimental results demonstrated the effectiveness of the residual transformer architecture and the hybrid loss function.
• The method outperformed several state-of-the-art polyp segmentation techniques.

Conclusions:

• The developed residual transformer-based model offers a promising solution for accurate and automated polyp segmentation.
• The novel hybrid loss function effectively addresses image-wise, pixel-wise, and regional inconsistencies in segmentation.
• This approach has the potential to significantly aid clinicians in early colorectal cancer detection and prevention.