DSI-Net: Deep Synergistic Interaction Network for Joint Classification and Segmentation With Endoscope Images

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Gastroenterology

Background:

• Automatic classification and segmentation of wireless capsule endoscope (WCE) images are crucial for diagnosing gastrointestinal diseases.
• Existing methods often treat these tasks independently, limiting diagnostic performance.
• There is a need for integrated approaches that leverage the complementary information between classification and segmentation.

Purpose of the Study:

• To propose a deep synergistic interaction network (DSI-Net) for the joint classification and segmentation of WCE images.
• To enhance the performance of computer-aided diagnosis systems for gastrointestinal diseases.
• To overcome the limitations of individually treated classification and segmentation tasks.

Main Methods:

• Developed a DSI-Net comprising classification (C-Branch), coarse segmentation (CS-Branch), and fine segmentation (FS-Branch).
• Introduced a lesion location mining (LLM) module in C-Branch to improve lesion highlighting using segmentation knowledge.
• Proposed a category-guided feature generation (CFG) module in FS-Branch to enhance segmentation using classification priors.
• Implemented a task interaction loss for mutual supervision and prediction consistency.

Main Results:

• DSI-Net demonstrated superior classification and segmentation performance compared to state-of-the-art methods on a public dataset.
• The synergistic interaction mechanism effectively integrated classification and segmentation tasks.
• LLM and CFG modules enhanced feature representation and task-specific performance.

Conclusions:

• The proposed DSI-Net effectively integrates classification and segmentation for WCE image analysis.
• Joint learning significantly improves diagnostic accuracy in computer-aided diagnosis systems.
• DSI-Net offers a promising approach for advancing WCE-based gastrointestinal disease detection.