Detection of Interpretable and Fine-Grained Brain Tumor Magnetic Resonance Imaging Based on Progressive Pruning:...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Brain tumors present significant detection challenges due to heterogeneity and similarity to normal tissue in MRI.
• Early and accurate detection is crucial for improving patient survival rates in central nervous system malignancies.

Purpose of the Study:

• To develop a high-performance brain tumor detection framework for MRI.
• To optimize the balance between detection accuracy, model efficiency, and interpretability.
• To enable accurate slice-level MRI tumor localization.

Main Methods:

• A novel convolution Prewitt-and-pooling-based preprocessing (CSPP) approach integrated into the YOLOv11 framework.
• Inclusion of a dynamic convolution-based C3k2 (DCC) module for enhanced feature capture.
• Integration of a channel prior convolutional attention (CPCA) module and progressive hybrid pruning strategy (PHPS) for model optimization and efficient inference.
• Utilized Eigen-class activation mapping (Eigen-CAM) for transparent prediction interpretation.

Main Results:

• The CDCP-YOLO framework demonstrated superior performance across three brain tumor MRI datasets.
• Significant improvements in mean average precision (mAP) were observed, with increases up to 19.5% on Roboflow dataset.
• Achieved substantial model efficiency gains, reducing floating-point operations (GFLOPs) by 47.7% on average across datasets.

Conclusions:

• The CDCP-YOLO framework offers an optimal balance of accuracy, efficiency, and interpretability for brain tumor detection.
• Provides a lightweight and reliable solution for slice-level brain tumor localization in MRI.
• Enhances the potential for improved clinical decision-making in neuro-oncology.