Abstract

Info

Author(s):: BAI Jingyi1; 2; WU Yirong1; 2; 3; LI Xiaolong1; 2; SUN Shuifa1; 2; 4; 1. Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering/College of Computer and Information Technology, China Three Gorges University, Yichang 443002, China; 2. Yichang Key Laboratory of Intelligent Medicine/College of Computer and Information Technology, China Three Gorges University, Yichang 443002, China; 3. Institute of Advanced Studies in Humanities and Social Sciences, Beijing Normal University, Zhuhai 519087, China; 4. School of Information Science and Technology, Hangzhou Normal University, Hangzhou 311121, China

Keywords:: brain tumor; YOLOv7; partial convolution; three-dimensional spatial attention; dynamic attention

Abstract:: Brain MRI data is characterized by large volumes and susceptibility to noise and artifacts, which pose significant challenges of improving the speed and accuracy of brain tumor detection and analysis due to the tumors’ diverse types, shapes, and boundaries that are both similar and highly variable. Therefore, a series of improvements based on YOLOv7 algorithm are proposed for enhancing detection precision and speed: (1) employing partial convolution during feature extraction to reduce the model’s parameters and improve overall detection speed; (2) in light of the complex variability of brain tumors, introducing a three-dimensional spatial attention mechanism during feature extraction to enhance the model’s focus on critical image features; (3) replacing the original IoU loss function with WIoU to increase the attention to mediumquality anchor boxes during bounding box regression for further improving detection accuracy. Experiments conducted on two public brain tumor datasets, Brain_Tumor and Glioma_of_test, show that the improved model achieves mAP of 96.9% and 92.8%, which are 1.4% and 2.4% higher than the original YOLOv7 model, and the frames per second reach 162.7 and 158.1, showing improvements of 6.4 and 18.2, respectively. These enhancements enable more effective detection of brain tumors in MRI images.