An enhanced YOLOv8 framework for MRI brain tumor detection: Integrating dense connectivity and dynamic attention mechanisms

This paper explores core obstacles in the application of object detection models to medical imaging, focusing on issues stemming from scale variation and fuzzy boundaries in MRI brain tumor analysis. We propose an improved detection framework that integrates a lightweight efficient backbone with a high-resolution detection strategy. Research centers on two core directions: optimizing internal feature extraction via dense connectivity and dynamic convolution, and enhancing cross-scale information fusion. Based on the You Only Look Once version 8 (YOLOv8) architecture, we develop a paradigm that replaces the standard C2f module with a fused CSP-DenseNet and KernelWarehouse structure. Subsequently, we introduce a Generalized Feature Pyramid Network (GFPN) embedded with Bi-directional Routing Attention (BRA) to capture long-range dependencies. Furthermore, the framework expands the detection logic to include a fourth highresolution head specifically for micro-tumor identification. Empirical validation on brain tumor datasets demonstrates significant improvements over existing baselines. Major breakthroughs are achieved in small object recall and mean Average Precision (mAP), with the proposed model achieving a 4.7% increase in [email protected] and an 11.2% improvement in micro-lesion recall while maintaining real-time performance. Consequently, this research offers a novel perspective for clinical early diagnosis systems.