TransConver: transformer and convolution parallel network for developing automatic brain tumor segmentation in MRI...

Area of Science:

• Medical imaging
• Artificial intelligence
• Biomedical engineering

Background:

• Medical image segmentation is crucial for computer-aided diagnosis (CAD).
• Convolutional Neural Networks (CNNs) excel at local feature extraction, while Transformers are adept at global representation.
• Effectively combining CNNs and Transformers for medical image segmentation remains a challenge due to semantic differences.

Purpose of the Study:

• To propose TransConver, a U-shaped network integrating convolution and transformer for accurate brain tumor segmentation in MRI.
• To address the semantic gap between local and global features in medical image segmentation.

Main Methods:

• Developed TransConver, a U-Net architecture utilizing a parallel TC-Inception module for simultaneous local (convolution) and global (transformer) feature extraction.
• Introduced a Cross-Attention Fusion with Global and Local features (CAFGL) mechanism within TC-Inception.
• Implemented a Skip Connection with Cross-Attention Fusion (SCCAF) to enhance encoder-decoder feature fusion.
• Designed both 2D and 3D versions of TransConver for diverse segmentation tasks.

Main Results:

• TransConver achieved superior performance on MICCAI BraTS datasets.
• Attained average Dice scores of 83.72% on BraTS2019 and 86.32% on BraTS2018.
• Demonstrated the model's effectiveness in accurately segmenting brain tumors.

Conclusions:

• TransConver effectively combines local and global features for improved brain tumor segmentation accuracy.
• The TC-Inception module successfully extracts both fine-grained local details and semantic global information.
• The proposed network architecture offers a significant advancement in medical image segmentation for CAD systems.