Seam-Stress: A Weakly Supervised Framework for Multiple Types of Interstitial Lung Disease Segmentation In Chest Ct

We propose Self-supervised Equivariant Attention Mechanism with up-Sampling, adapTive thREsholding, and augmented loSSes (SEAM-STRESS) for weakly supervised abnormality segmentation in chest CT images, covering a wide range of lung abnormality patterns. We introduce a novel point-level loss that allows the utilization of sparse annotations as a weak supervisory signal, outperforming models trained only with image-level labels. Furthermore, we introduce a post-processing adaptive background thresholding strategy to further improve the segmentation masks. Experiments on our internal dataset show the effectiveness of our framework in localizing and segmenting chest CT abnormalities.