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

IntroductionBrain tumor lesions in MRI images are characterized by low contrast, indistinct boundaries, and irregular morphology, posing substantial challenges for automated detection. RT-DETR provides an end-to-end, NMS-free detection paradigm, but its original design is optimized for natural images and exhibits three domain-specific limitations in brain tumor MRI detection: insufficient local texture perception of low-contrast lesion boundaries, neglect of feature magnitude differences in linear attention, and weak modeling of spatial continuity for morphologically irregular tumors. This study proposes CMRA-DETR (CSP-MambaOut with Retention and Magnitude-Aware Attention for Real-Time Detection Transformer), a lightweight yet high-accuracy detection framework built upon RT-DETR-R18, to address these limitations through targeted architectural adaptations.MethodsCMRA-DETR introduces three improvements: (1) a CSP-MambaOut backbone that enhances local texture perception of low-contrast lesion boundaries via gated feature selection; (2) an AIFI-MALA module that introduces magnitude-aware linear attention to correct the distributional smoothing deficiency of standard linear attention; and (3) a RetBlockC3 module that incorporates a Manhattan distance decay-based spatial retention mechanism to improve modeling of spatial continuity for morphologically irregular tumors. The model was trained and evaluated on an internal independent test set of 5,731 MRI images covering four categories (no tumor, meningioma, glioma, and pituitary adenoma), and further assessed on an external independent test set from the BRISC dataset without fine-tuning. Its performance was benchmarked against Faster R-CNN, YOLOv5n, YOLOv8n, YOLO11n, YOLO12n, and the RT-DETR-R18 baseline.Results and discussionOn the internal independent test set, CMRA-DETR achieved achieves p = 95.5%, R = 95.7%, mAP@50 = 97.9%, and mAP@50-95 = 82.6%, while reducing parameters and GFLOPs by 37.7% and 30.9%, respectively, relative to the baseline, achieving competitive or superior performance compared to the evaluated baseline models under the specified experimental conditions. On the external independent test set from the BRISC dataset, the model achieved mAP@50 = 96.6% and mAP@50-95 = 79.8% without fine-tuning, demonstrating robust cross-dataset generalization. These findings indicate that CMRA-DETR achieves a favorable balance among detection accuracy, model lightweightness, and inference efficiency, demonstrating practical potential for AI-assisted brain tumor detection on resource-constrained clinical devices.