Superpixel-based Context Restoration for Self-supervised Pancreas Segmentation from CT scans

Automatic segmentation of the pancreas can help research in pancreatic cancer and other pancreatic diseases. Quantitative measures which are extracted from the pancreas based on CT imaging provide valuable biomarkers for tracking the progression of various endocrine and exocrine diseases. In recent years, deep learning has proven to be a powerful tool for pancreas segmentation. However, deep learning models in medical image analysis suffer from data scarcity: the lack of annotated data poses a significant drawback in developing new models. One possible solution is self-supervised learning which comprises of an unsupervised pre-training stage followed by a down-stream supervised learning stage. This paper presents a superpixel based approach to construct a pretraining task for self-supervised learning for pancreas segmentation. We corrupt the CT images segmented with superpixels by replacing random segments with intensity values randomly sampled from the image. The weights learnt when the model reconstructs the image is used to initialize network weights in the downstream segmentation task. We used 59 CT scans from the AbdomenCT-1k dataset for pre-training and 82 CT scans from NIH pancreas-CT dataset for the downstream segmentation task. We achieved an increase in performance with our approach compared to the randomly initalized weights, as contextual image features are learnt via this context restoration. Moreover, our approach outperforms existing context restoration approaches using patch based methods