BTA-DETR: a multi-domain cooperative perception model for brain tumor detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Brain tumor MRI detection faces challenges due to lesion heterogeneity, blurred boundaries, and noise.
• Existing models struggle to accurately capture diverse lesion morphologies and fine-grained features.

Purpose of the Study:

• To develop an advanced multi-domain cooperative perception model for improved brain tumor MRI detection.
• To enhance the model's ability to handle complex lesion characteristics and improve localization accuracy.

Main Methods:

• Proposed BTA-DETR (Brain Tumor Aware-DETR) model incorporating a Content-Aware Fusion Unit (CAFU) for adaptive feature re-weighting.
• Introduced Learnable Temperature Attention (LTA) for precise localization of blurred boundaries.
• Developed a Frequency-Spatial Cooperative Module (FSCM) to integrate frequency and spatial domain information for enhanced feature representation.

Main Results:

• BTA-DETR improved mAP50 by 4.36 percentage points compared to the baseline RT-DETR on the Roboflow dataset.
• The model demonstrated a reduction of 0.82M parameters and approximately 9.1% in computational cost.
• BTA-DETR achieved higher detection metrics than the vanilla RT-DETR on external datasets (Kaggle BrainTumor, BraTS 2021 T1ce).

Conclusions:

• BTA-DETR offers a significant advancement in brain tumor MRI detection, effectively addressing challenges of heterogeneity and blurred boundaries.
• The proposed model provides a more efficient and accurate solution for brain tumor detection, outperforming existing methods.