BrainYOLO-MCA: An Improved YOLOv11 with Multi-Scale Channel Attention for Brain Tumor Detection

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neurosurgery and Oncology

Background:

• Brain tumor detection in MRI is vital for early diagnosis and treatment planning.
• Challenges include blurred boundaries, tissue similarity, and small lesions, hindering accurate detection.
• Existing methods often struggle with these complexities, necessitating advanced solutions.

Purpose of the Study:

• To propose a novel brain tumor detection framework, BrainYOLO-MCA, to overcome current detection limitations.
• To enhance the recognition of indistinct tumor boundaries and improve sensitivity to tiny early-stage lesions.
• To provide a robust and accurate tool for clinical decision-making in neuro-oncology.

Main Methods:

• BrainYOLO-MCA is an enhanced YOLOv11 architecture incorporating a Multi-scale Channel Attention (MCA) mechanism.
• Key optimizations include dynamic upsampling (DySample) for spatial details and a small-object detection head.
• A sparse self-attention mechanism was integrated to strengthen global contextual features.

Main Results:

• BrainYOLO-MCA achieved high precision (up to 98.6%) and mAP50 (up to 97.5%) across three datasets (Br35H, MBrT, Figshare).
• The model outperformed baseline YOLOv11 and other mainstream YOLO-based models.
• It maintained a lightweight architecture with only 3.1M parameters.

Conclusions:

• BrainYOLO-MCA significantly improves the detection of brain tumors, especially those with blurred boundaries and small sizes.
• The MCA mechanism and feature fusion strategies enhance robustness and detection accuracy.
• The framework supports more accurate and reliable brain tumor detection for clinical decision-making and early diagnosis.