Abstract

Intracranial hemorrhage (ICH) and unruptured intracranial aneurysm (UIA) are two important cerebrovascular diseases that require prompt and precise diagnosis for effective treatment and improved survival rates. While deep learning (DL) techniques have emerged as the leading approach for medical image analysis and processing, the most commonly employed supervised learning often requires large, high-quality annotated datasets that can be costly to obtain, particularly for pixel/voxel-wise image segmentation. To address this challenge and meet the need in cerebrovascular care, we proposed and validated three novel weakly supervised segmentation methods for ICH using categorical labels and for UIA with coarse image segmentation. For ICH, we first introduced a framework to segment the lesion based on a hierarchical combination of self-attention maps obtained from a Swin transformer, which was trained only for ICH detection, achieving a Dice score of 0.407. Subsequently, by employing novel head-wise gradient-weighing of self-attention maps in the same setup, we further improved the mean Dice score to 0.444 for ICH segmentation. Our method that only relies on categorical labels showed comparable performance against popular fully supervised methods, such as UNet and Swin-UNETR. Finally, for UIA segmentation, we achieved a Dice score of 0.68 and a 95% Hausdorff distance of ~0.95 mm by proposing a new 3D focal modulation UNet, called FocalSegNet. This novel DL architecture was trained with coarse manual segmentation, providing an initial segmentation of aneurysms, which was then refined using dense conditional random field (CRF) post-processing. Our proposed methods explored new avenues using weak labels to mitigate a key bottleneck in medical DL with excellent performance and showcased their promising potential in addressing challenging medical image segmentation tasks.