Complementary learning for Deep Medical Image Segmentation and Anomaly Detection

Image segmentation is a fundamental step in many biomedical image analysis applications toaccurately and coherently separate out an region of interest for further analysis. However often times additional information and features which are not directly related to the object of interest is discarded. This dissertation explores ways to leverage those complementary information in order to enhance the segmentation of an object of interest. There are three major contributions; Firstly, the idea of complementary learning is formulated as a simple to use loss function for use with deep learning frameworks. Secondly, we contribute a model, Comp-Net, for supervised learning which enhances an existing segmentation algorithm to be more tolerant to unseen deviation during test set uncharacteristic to the training set. We show the robustness on two applications, skull stripping and liver lesion segmentation. Finally, we propose a new way to perform Unsupervised Anomaly Segmentation via selective two cuts with one cut chosen by the user as a reference distribution, outperforming existing state of the art methods on MS-SEG2015 dataset and show promising performance on Brats2019, LiTS Liver Lesion and a privately curated brain tumor dataset. The two cut allows to move away from reconstruction dependent anomaly/novelty segmentation and moves the focus back to the anomaly itself.