Two-Stage Approach to Intracranial Hemorrhage Segmentation From Head CT Images

Intracranial hemorrhage (ICH) is an emergency and a potentially life-threatening condition. Automated segmentation of ICH from head CT images can provide clinicians with volumetric measures that can be used for diagnosis and decision support for treatment procedures. Existing solutions typically involve training deep learning models to perform segmentation directly on the whole CT image. However, datasets with segmentation masks are typically very small in comparison with datasets with bounding boxes. Thus, we propose a two-stage approach that utilizes both bounding boxes and segmentation masks to help improve segmentation performance. In the first stage, ICH regions are detected and localized with bounding boxes surrounding the lesion by using a supervised YOLOv5 object detector. In the second stage, the localized ICH foreground is automatically segmented using TransDeepLab, an attention-based transformer network. Although we utilize both ground-truth bounding boxes and segmentation masks, different datasets can be used to train each stage. There is no requirement for pairing up bounding boxes and segmentation masks to train the model. Since bounding box annotations are available in larger quantities than segmentation masks, our approach allows these large datasets of bounding boxes to be used to improve ICH segmentation performance. On our dataset of segmentation masks, we demonstrated that our proposed two-stage YOLOv5+ TransDeepLab model outperformed segmentation methods such as SegResNet by 8% in terms of Dice score. Given ground truth bounding boxes, a Dice score of 0.769 is achieved, outperforming state-of-the-art methods such as nnU-Net. In sum, our proposed two-stage approach produces more accurate binary segmentation of ICH for neuroradiologists and these improved measurements could potentially aid their clinical decision-making process.