Multi-Scale Intracranial Hemorrhage Detection in CT scans Using Cross-Stage Partial Fusion and Hierarchical Feature Learning

Intracranial hemorrhage (ICH) detection from CT scans poses two major challenges: (1) Variability in hemorrhage size and location, which requires a model capable of detecting small, medium, and large-scale ICH instances, and (2) Efficient feature extraction, ensuring accurate detection while maintaining computational efficiency. To address these challenges, this study proposes a deep learning-based architecture for automated ICH detection using CT scan images from the CQ500 dataset. The model leverages Mosaic Data Augmentation to enhance training and a customized C2F (Cross-Stage Partial Fusion) module to improve feature extraction efficiency across different scales. The proposed model contains three main components: “the Backbone, Neck, and Head”. The C2F module, integrated into both the Backbone and Neck, enhances multi-scale feature learning by maintaining rich spatial information and improving gradient flow. The Neck incorporates Upsample layers to refine extracted features before passing them to the Head. The Detect module, embedded in the Head, performs hierarchical ICH predictions at small, medium, and large scales, ensuring comprehensive hemorrhage localization. The model was evaluated on the CQ500 dataset, achieving high performance metrics, with Precision: 0.945, Recall: 0.941, F1-score: 0.94, mAP@0.5: 0.974, and mAP@0.5:0.95: 0.822. These results demonstrate the model’s efficacy in accurately detecting ICH at multiple scales.