Research on Brain Tumor Detection in MRI Images Based on Improved YOLOv13 Algorithm

This study aims to improve the detection performance of brain tumor lesions in MRI images by proposing an improved YOLOv13 algorithm integrating switchable atrous convolution (SAConv), deformable attention (DAT), and lightweight dynamic upsampling (Dysample) module. Based on the YOLOv13 architecture consisting of the Backbone, Neck, and Head, we replaced some Convolutional (Conv) layers in the Backbone and Neck with SAConv, integrated the DAT mechanism at the end of the deep A2C2f module in the Backbone, and replaced all upsampling units in the Neck with enhanced Dysample module, ultimately constructing the YOLOv13-SAConv-DAT-Dysample algorithm. Training and validation were conducted using the public MRI brain tumor detection dataset from the Roboflow platform, and a comprehensive performance comparison was made with the original YOLOv13 algorithm. The results show that the mean average precision (mAP) of the improved algorithm under Intersection over Union (IoU) thresholds of 0.5, 0.75, and 0.5~0.95 reached 0.914, 0.783, and 0.667, respectively, which are comprehensively superior to the original algorithm (with corresponding metrics of 0.882, 0.766, and 0.648). The conclusion confirms that the YOLOv13-SAConv-DAT-Dysample algorithm can effectively realize the detection of brain tumor lesions in MRI images and has application potential for assisting clinical diagnosis.