A lightweight improved YOLOv11 framework for intracranial hemorrhage detection

Timely and accurate detection of intracranial hemorrhage in CT images is essential for emergency intervention and for reducing disability and mortality. Although deep learning-based object detectors have shown considerable promise for computer-aided intracranial hemorrhage detection, their performance in CT-based detection settings remains constrained by subtle lesion characteristics, insufficient multi-scale feature representation, class imbalance, and complex background interference. In addition, many existing models still struggle to balance detection precision with computational efficiency, which limits their suitability for real-time diagnosis. To overcome these limitations, we propose a lightweight YOLOv11-based framework for intracranial hemorrhage detection. It improves feature extraction, feature fusion, and lesion localization while maintaining practical inference efficiency. Within this framework, RepStem is adopted to replace the original downsampling convolution for stronger early lesion representation, the FDPN-DASI joint feature optimization module is introduced to improve multi-scale hematoma aggregation and suppress irrelevant background responses, and the original C2PSA block is replaced with cascaded C2CGA attention to better capture small hemorrhages and ambiguous boundaries. Experiments on the BHX dataset show that the proposed method outperforms the YOLOv11 baseline and achieves competitive overall performance among the compared detectors. Compared with YOLOv11, our method improves precision and recall by 3.9% and 3.4%, respectively, while increasing [email protected] and [email protected]:0.95 by 3.7% and 4.6%. Despite these precision gains, the model remains lightweight, containing only 2.65 million parameters and achieving 864.86 FPS, suggesting its potential for CT-based computer-aided detection of intracranial hemorrhage.