A teacher-student framework for liver and tumor segmentation under mixed supervision from abdominal CT scans

Liver and tumor segmentation from abdominal CT scans and an important step towards computer-assisted diagnosis or treatment planning for various hepatic diseases. Training convolutional neural networks for image segmentation demands a large number of pixel-wise labels which are inefficient to acquire. In order to leverage massive weak annotations, we developed a teacher-student framework using both pixel annotated dataset (strong dataset) and bounding box annotated dataset (weak dataset). A teacher annotator transfers the knowledge from the strong dataset to the weak one by refining its bounding box labels into pseudo pixel-wise labels. Motivated by the spatial layout of organ and tumor, we proposed a hierarchical organ-to-lesion (O2L) attention module to regularize the teacher annotator trained on the strong dataset. A student segmentor is trained with the mix of strong and refined weak datasets. A localization branch in the student network aggregates deep features to predict positions of organ and lesion, improving the segmentation of small objects. A comparative study with state-of-the-art methods demonstrates the proposed method strikes the balance between model performance and annotation efficiency. This model shows robustness to the quality of bounding box annotations. The model is also validated on kidney and tumor segmentation.