TDFormer: Top-Down Token Generation for 3D Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Accurate medical image segmentation is crucial for effective treatment planning.
• Current transformer-based methods use fixed grids, leading to inefficient processing of less important image regions.
• Unequal importance of image tokens necessitates adaptive computational resource allocation.

Purpose of the Study:

• To introduce a novel transformer-based segmentation framework, Top-Down Transformer (TDFormer).
• To develop a spatially adaptive token generation scheme for efficient medical image segmentation.
• To enhance computational focus on critical image areas, such as tumors, for higher resolution processing.

Main Methods:

• Proposed TDFormer framework with a spatially adaptive token generation scheme.
• Incorporated three key components: attentiveness calculation, token splitting, and token fusion.
• Gradual fusion of redundant background tokens to concentrate on salient regions.

Main Results:

• TDFormer demonstrates superior performance compared to state-of-the-art methods.
• Achieved high accuracy on publicly available datasets: BTCV Challenge, LiTS, and BraTS 2020.
• Experimental analysis confirmed the robustness and effectiveness of each component within TDFormer.

Conclusions:

• TDFormer offers a robust and effective solution for medical image segmentation.
• The spatially adaptive token generation scheme significantly improves computational efficiency.
• This approach enables more focused processing of critical details in medical images.