HEIDet: Edge-guided hierarchical network for lightweight and high-precision brain tumor detection

Early screening of brain tumors is critical for clinical diagnosis, treatment, and prognosis. However, diagnose relying on time-consuming and highly subjective manual analysis of medical images are prone to missed or misdiagnosed cases, ultimately hindering their clinical scalability. Meanwhile, due to tumor heterogeneity, low boundary contrast, and the prevalence of sub-centimeter lesions in MRI images, deep learning-based brain tumor detection remains a formidable challenge in the field of medical imaging. To address these issues, we propose a medical-specific optimized brain tumor detection framework called the Hierarchical Edge Integration Detector (HEIDet), which introduces two clinically motivated innovative modules: Edge Integration Multi-scale Selection (EIMS) module and Hierarchical Re-parameterized Interactive Fusion (HRIF) module. Finally, HEIDet respectively achieves 95.8% m A P 0 . 5 and 74.1% m A P 0 . 5 : 0 . 95 on the Br35H dataset, and for multi-class detection on MBTBB dataset, it attains 97.3% m A P 0 . 5 and 80.6% m A P 0 . 5 : 0 . 95 . Results on both datasets demonstrate that HEIDet out-performs other state-of-the-art models, highlighting its effectiveness in early brain tumor screening and computer aided diagnosis. The code is available at: https://github.com/VincentYuuuuuu/HEIDet . • Proposes HEIDet, a lightweight framework for efficient brain tumor detection. • Designs EIMS module to enhance feature extraction using edge information. • Develops HRIF module for adaptive fusion of multi-scale features. • Achieves SOTA results with only 3.86M parameters, validated by radiologist. • Provides visualization and ROC analysis for intuitive model interpretation.