Detection of Interpretable and Fine-Grained Brain Tumor Magnetic Resonance Imaging Based on Progressive Pruning: Machine Learning Model Development and Validation Study.

Background: Brain tumor is one of the most malignant diseases of the central nervous system, and early accurate detection is of great significance for improving patient survival rate. However, the heterogeneity of brain tumors in terms of morphology, size, and location on magnetic resonance imaging (MRI) image, as well as their similarity to surrounding normal brain tissue, poses significant challenges for tumor detection. Objective: This study aims to develop a high-performance brain tumor detection framework that integrates feature enhancement, channel attention, and progressive pruning, achieving an optimal balance between detection accuracy, model efficiency, and interpretability for slice-level MRI tumor localization tasks. Methods: This paper proposes a convolution Prewitt-and-pooling-based preprocessing (CSPP) approach, based on the "you only look once" version 11 (YOLOv11) framework, which highlights important structural detail more effectively than traditional statistics. A dynamic convolution-based C3k2 (DCC) module was integrated to more efficiently capture both local and global features. A channel prior convolutional attention (CPCA) module was introduced before the detection head, enabling the network to specifically focus on information-rich channels and key spatial regions. Through a progressive hybrid pruning strategy (PHPS), the model was optimized for efficient inference. Furthermore, Eigen-class activation mapping (Eigen-CAM) was used to interpret the prediction result, making them more transparent. Results: Extensive experiments on 3 brain tumor MRI datasets demonstrated the superior performance of CDCP-YOLO (CSPP-DCC-CPCA-PHPS-YOLO). On Br35H, the mean average precision (mAP) at an intersection-over-union (IoU) threshold of 0.5 (mAP0.5) increased by 2.6%, average mAP over several IoU thresholds (0.50-0.95; mAP0.5:0.95) increased by 5.9%, and number of floating-point operations (×10⁹; GFLOPs) decreased by 47.7%. On Roboflow, mAP0.5 increased by 19.5%, mAP0.5:0.95 increased by 7.7%, and GFLOPs decreased by 47.7%. On Capstone, mAP0.5 increased by 6.9%, mAP0.5:0.95 increased by 5.8%, and GFLOPs decreased by 47.7%. Conclusions: The proposed CDCP-YOLO framework achieves an optimal balance between accuracy, efficiency, and interpretability, providing a lightweight and reliable solution for slice-level brain tumor detection in MRI images.