Anatomy-aware transformer-based model for precise rectal cancer detection and localization in MRI scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Rectal cancer diagnosis relies on MRI, but detection is challenging due to image complexity and localization difficulties.
• Transformer models show promise in object detection but face limitations in medical imaging due to scarce data.

Purpose of the Study:

• To enhance the accuracy of rectal cancer detection in MRI scans using a novel transformer-based approach.
• To address the challenge of detecting small rectal cancers by integrating anatomical information and global context.

Main Methods:

• Proposed the Spatially Prioritized Detection Transformer (SP DETR) with a Spatially Prioritized (SP) Decoder to focus on regions of interest (ROI) using anatomical maps.
• Introduced the Global Context-Guided Feature Fusion Module (GCGFF) with a transformer encoder and Globally-Guided Semantic Fusion Block (GGSF) to improve feature representation.
• Utilized SP cross-attention to refine anchor box offset learning.

Main Results:

• The SP DETR model significantly improved rectal cancer detection accuracy in MRI scans.
• The model demonstrated particular effectiveness in detecting small rectal cancers, a known challenge in the field.
• Integration of anatomical priors and global context enhanced the transformer model's performance.

Conclusions:

• The SP DETR model offers a promising advancement for accurate rectal cancer diagnosis from MRI scans.
• Incorporating anatomical knowledge and global context into transformer architectures is effective for medical image analysis.
• This approach holds potential for improving clinical outcomes in rectal cancer management.