Efficient 3D Biomedical Image Segmentation by Parallelly Multiscale Transformer-CNN Aggregation Network.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Deep Learning for Segmentation

Background:

• Accurate 3D biomedical image segmentation is crucial for clinical diagnosis, surgery, and prognosis.
• Existing deep learning models face challenges in simultaneously capturing global and local image features for precise segmentation.

Purpose of the Study:

• To develop an advanced segmentation solution, the mixed parallel shunted transformer (MPSTrans), for improved 3D biomedical image analysis.
• To address the limitations of current methods in comprehensive feature capture and multiscale synchronization.

Main Methods:

• Introduced MPSTrans, a novel deep learning model featuring 3D-MPST blocks within a U-form framework.
• Implemented deep supervision in the decoder for hierarchical representation learning.
• Evaluated performance on colon cancer, multi-organ, and multimodal datasets.

Main Results:

• MPSTrans demonstrated significant improvements in Dice Similarity Coefficient (DSC) and a reduction in Hausdorff Distance at 95% (HD95) on colon cancer data.
• Achieved a 56.7% reduction in computational load (GFLOPs).
• Outperformed mainstream methods like Swin UNETR, UNETR, nnU-Net, PHTrans, and 3D U-Net on public datasets (MSD, BCV, ACDC).

Conclusions:

• MPSTrans offers a robust and adaptable solution for 3D biomedical image segmentation, enhancing diagnostic capacity.
• The model's ability to capture comprehensive features and reduce computational load positions it as a state-of-the-art tool for medical imaging analysis.