TransDiffSeg: Transformer-Based Conditional Diffusion Segmentation Model for Abdominal Multi-Objective.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Traditional diffusion probabilistic models for medical image segmentation suffer from local inductive biases of convolutional operations, limiting long-term dependency modeling and segmentation accuracy.
• Transformers can overcome these limitations by removing local inductive biases, enhancing segmentation precision.

Purpose of the Study:

• To propose TransDiffSeg, a conditional diffusion segmentation model that integrates Transformer and convolution operations in parallel for finer granularity bias elimination.
• To enhance global semantic information and reduce Transformer's noise sensitivity through an adaptive feature fusion block.

Main Methods:

• Developed TransDiffSeg, a conditional diffusion segmentation model using a parallel combination of Transformer and convolution operations.
• Implemented an adaptive feature fusion block to merge semantic and noise features.
• Conducted experiments on the AMOS22 and BTCV datasets to evaluate performance.

Main Results:

• Eliminating local inductive bias at a finer granularity significantly improved segmentation performance in diffusion probabilistic models.
• The parallel integration of Transformer and convolution operations at a finer granularity outperformed existing nesting and stacking methods.
• Experimental results confirmed that finer granularity in bias elimination leads to better segmentation outcomes.

Conclusions:

• TransDiffSeg effectively addresses the limitations of traditional models by eliminating local inductive biases at a finer granularity.
• The proposed parallel integration strategy and adaptive feature fusion enhance medical image segmentation accuracy and robustness.
• Finer granularity in bias elimination is a critical factor for improving the performance of diffusion probabilistic models in medical image segmentation.