Capturing fine-grained spatial details in brain tumor MRI via multi-module feature enhancement

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Existing object detection models often face challenges in medical imaging, including limited feature fusion capabilities, constrained receptive fields, and the prevalent issue of class imbalance. To address these limitations, this paper proposes FDANet (Fine-Grained Detail Aggregation Network), a novel multi-module feature enhancement network built upon the YOLOv11 architecture. Specifically, FDANet integrates a Mixed Aggregation Network (MANet) to capture fine-grained spatial details of tumor regions. The FasterCGLU gating mechanism is employed to enhance lesion feature responses while suppressing background noise. Furthermore, a Wavelet Feature Upgrade (WFU) module is introduced to facilitate the adaptive fusion of multi-scale features. The synergy of these three components directs the model’s focus efficiently toward critical pathological regions, an efficacy substantiated by both heatmap visualizations and ablation studies. Datasets for this study were constructed by combining in-house brain tumor images with the public Alibaba Tianchi dataset. Extensive comparisons against mainstream detection models demonstrate that FDANet achieves precise identification of pituitary adenomas, gliomas, and meningiomas, attaining an mAP 50 of 0.961. This marks a 6.0 percentage point improvement over the baseline YOLOv11, all while maintaining a low parameter count and computational complexity. This study presents a multi-module feature enhancement framework for brain tumor MRI detection, demonstrating promising performance on a single-center 2D dataset. Further prospective validation on multi-center, volumetric data is warranted before clinical translation.