Abstract: Objective:This study aimed to enhance the detection efficiency of brain tumor images by optimizing the lightweight real-time object detection model YOLOv7-tiny, proposing an improved framework tailored for brain tumor detection tasks. Methods:Experiments were conducted using a benchmark dataset comprising 1,000 brain tumor images acquired using magnetic resonance imaging (MRI). The InceptionNeXt module was introduced into the backbone of the YOLO network to expand the receptive field. Multiscale feature fusion was employed to enhance the model’s ability to characterize tumors of varying sizes and shapes. Furthermore, a decoupled head structure was incorporated, adopting an independent branch optimization strategy to mitigate interference from complex backgrounds. This improved localization accuracy for small and boundary-blurred tumors, accelerated model convergence, and enhanced detection performance. Results:The improved model achieved a notable performance gain compared to the baseline YOLOv7-tiny. Mean average precision (mAP)@0.5 increased from 85.6% to 88.2%, and mAP@0.5:0.95 increased from 63.8% to 65.4%, corresponding to improvements of 2.6% and 1.6%, respectively. The experimental results demonstrate that the enhanced model can identify brain tumors more accurately and exhibits a relatively precise detection performance. Conclusion:The improved model was substantially more accurate than the original model, which validates its feasibility for accurate brain tumor detection.