Abstract
Segmenting abnormalities is a leading problem in medical imaging. Using machine learning for segmentation generally requires manually annotated segmentations, demanding extensive time and resources from radiologists. We propose a weakly supervised approach that utilizes binary image-level labels, which are much simpler to acquire, rather than manual annotations to segment brain tumors on magnetic resonance images. The proposed method generates healthy variants of cancerous images for use as priors when training the segmentation model. However, using weakly supervised segmentations for downstream tasks such as classification can be challenging due to occasional unreliable segmentations. To address this, we propose using the generated non-cancerous variants to identify the most effective segmentations without requiring ground truths. Our proposed method generates segmentations that achieve Dice coefficients of 79.27% on the Multimodal Brain Tumor Segmentation (BraTS) 2020 dataset and 73.58% on an internal dataset of pediatric low-grade glioma (pLGG), which increase to 88.69% and 80.29%, respectively, when removing suboptimal segmentations identified using the proposed method. Using the segmentations for tumor classification results with Area Under the Characteristic Operating Curve (AUC) of 93.54% and 83.74% on the BraTS and pLGG datasets, respectively. These are comparable to using manual annotations which achieve AUCs of 95.80% and 83.03% on the BraTS and pLGG datasets, respectively.