Utilizing shallow features and spatial context for weakly supervised intracerebral hemorrhage segmentation.

Background: In deep learning-assisted intracerebral hemorrhage (ICH) diagnosis, because weak image-level labels cannot provide supervisory information about the target location, weakly-supervised semantic segmentation (WSSS) methods are relatively limited. Therefore, we developed a novel method for improving the ICH segmentation results from weak image-level labels. Methods: This paper proposes the Shallow-Feature class activation map (CAM) module, which utilizes fine-grained information from the shallow feature maps of convolutional neural networks (CNNs) to generate CAM for accurate target localization and contour. Then, the Spatial Context Aware (SCA) module utilizes the spatial context information in computed tomography (CT) images to further complete the hemorrhage sites that the CAM of the current slice failed to locate. Finally, we binarize the CAM based on the selected threshold to generate pseudo-segmentation masks. Additionally, we used two publicly available ICH segmentation datasets, the Brain Hemorrhage Segmentation Dataset (BHSD) and the CT Images for Intracranial Hemorrhage Detection and Segmentation Dataset (BCIHM), to verify the efficiency of our proposed method. Results: Our results showed that our proposed method is effective in improving the accuracy of ICH segmentation, with the mean Intersection over Union (mIoU) increasing from 52.5% to 69.8% in BHSD and 50.1% to 68.9% in BCIHM. The segmentation results of ICH generated by our method were superior to other WSSS methods, with a mIoU of 69.8% and 68.9%, correct localization of 48.1% and 48.9%, missed localization of 51.9% and 51.1%, false positive localization of 49.8% and 51.2%, reached 88% and 86% of the fully supervised U-Net performance, in BHSD and BCIHM, respectively. Conclusions: Our proposed method can better match the location and contours of hemorrhage points, significantly reducing missed localization and false positive localization, thus effectively reducing the workload of professional radiologists in creating pixel-level datasets.