TransRender: a transformer-based boundary rendering segmentation network for stroke lesions.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computational Neuroscience

Background:

• Vision transformers excel at capturing global features in medical image segmentation.
• Current methods struggle with accurate lesion segmentation due to complex brain structures and similar tissue/lesion appearances.
• Existing decoders often overlook high-frequency boundary details, focusing instead on regional features.

Purpose of the Study:

• To develop an effective method for precise lesion boundary rendering in medical images.
• To address the limitations of existing segmentation techniques in accurately delineating lesion edges.
• To improve the accuracy and efficiency of stroke lesion segmentation.

Main Methods:

• Proposed TransRender, a novel method utilizing point-based rendering for lesion boundary feature computation.
• Employed a transformer-based encoder to capture global information.
• Integrated renders to map encoded features to original spatial resolution, combining global and local information, with point-based supervision for refinement.

Main Results:

• TransRender adaptively selects important points to compute boundary features, enhancing boundary representation.
• Experiments on stroke lesion segmentation datasets demonstrated the method's efficiency.
• Achieved high accuracy and low computational complexity in automatic stroke lesion segmentation.

Conclusions:

• TransRender effectively refines uncertainty regions through continuous point generation and supervision.
• The proposed method offers a significant advancement in accurate and efficient medical image segmentation, particularly for brain lesions.
• Demonstrated superior performance over existing methods in segmenting stroke lesions.