Unleashing Diffusion and State Space Models for Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Medical Imaging Analysis

Background:

• Current medical image segmentation models struggle with robustness on unseen data.
• Identifying rare or novel tumor types requires advanced segmentation capabilities.

Purpose of the Study:

• To develop a robust segmentation framework (DSM) for identifying unseen tumor categories.
• To enhance the model's ability to generalize across diverse medical imaging scenarios.

Main Methods:

• Utilized a novel framework combining diffusion and state space models (DSM).
• Employed dual object query sets within modified attention decoders for enhanced classification.
• Incorporated object-aware feature grouping for organ queries and diffusion-based prompts for tumor queries.
• Integrated diffusion-guided feature fusion and CLIP text embeddings for improved semantic segmentation and linguistic transfer.

Main Results:

• DSM demonstrated superior performance in segmenting unseen tumor categories.
• Achieved significant improvements in out-of-distribution detection metrics: 0.1962 in mean AUROC, 0.2675 in mean FPR95, and 0.1736 in mean DSC.
• Showcased enhanced robustness in multi-label tasks and diverse segmentation scenarios.

Conclusions:

• DSM offers a robust solution for medical image segmentation, particularly for unseen and rare tumor types.
• The framework significantly advances the capabilities of AI in medical diagnostics and analysis.