CMRA-DETR: a lightweight and high-accuracy detection framework for MRI-based brain tumor identification

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Deep learning for medical diagnosis

Background:

• Brain tumor detection in MRI is challenging due to low contrast, indistinct boundaries, and irregular shapes.
• Existing real-time detection transformers (RT-DETR) optimized for natural images struggle with brain tumor MRI specifics.
• Limitations include poor local texture perception, neglect of feature magnitude, and weak spatial continuity modeling.

Purpose of the Study:

• To develop a lightweight, high-accuracy detection framework for brain tumors in MRI.
• To address the domain-specific limitations of RT-DETR in brain tumor detection.
• To propose CMRA-DETR (CSP-MambaOut with Retention and Magnitude-Aware Attention for Real-Time Detection Transformer) for enhanced performance.

Main Methods:

• Introduced a CSP-MambaOut backbone for improved local texture perception.
• Incorporated an AIFI-MALA module with magnitude-aware linear attention to address distributional smoothing.
• Utilized a RetBlockC3 module with spatial retention for irregular tumor morphology.
• Trained and evaluated on 5,731 internal MRI images and the external BRISC dataset.

Main Results:

• CMRA-DETR achieved 95.5% precision and 95.7% recall on the internal test set.
• Achieved mAP@50 of 97.9% and mAP@50-95 of 82.6% on the internal test set.
• Reduced parameters by 37.7% and GFLOPs by 30.9% compared to the baseline.
• Demonstrated strong cross-dataset generalization on the BRISC dataset (mAP@50 = 96.6%).

Conclusions:

• CMRA-DETR offers a superior balance of accuracy, lightweight design, and inference efficiency for brain tumor detection.
• The model shows significant potential for AI-assisted diagnosis in resource-constrained clinical settings.
• Architectural adaptations effectively address the challenges of brain tumor MRI analysis.