Multi-modal medical image Segmentation using convolutional neural networks with foreground background reciprocity

Abstract

Precise medical image segmentation is a crucial step for proper isolation of target re-gions, such as an organ or lesion for accurate medical diagnosis, prognosis and certain medical procedures. Despite making significant progress in the field of computer aided segmentation, existing algorithms still fail to generalize well to diverse modalities of medical images. These conventional algorithms fail to adequately account for diverse challenges posed by different segmentation tasks. In this thesis, the above limitation is addressed in a series of two studies. In the first study, we introduce an attention guided bipolar refinement-based segmentation network, which employs a complementary at-tention scheme between a pair of positive and negative refinement modules placed on top of two encoder structures for generating refined feature references for the decoder stage in a supervised manner. A four-way feature shifting operation is introduced in conjunction with a set of dilated convolutional layers so that it considers the spatial re-lationships across a wider footprint leading to better contextual feature extraction. We also formulate a novel Foreground-to-Background Ratio (FBR) index to highlight the differences in signal power between target region and background due to the refinement brought about by the refinement modules. In the second part of the thesis, we present a methodology which contrasts the general theme of contemporary literature on medical image segmentation. We demonstrate that considering the background tissue segmenta-tion task alongside the main foreground task can improve overall segmentation perfor-mance when considered from a general medical image segmentation perspective. We, therefore, propose a novel DL framework that ties together two streams (foreground and background) through an image reconstruction task. A boxed Mean Squared Error loss is proposed to complement the dice losses from both streams. We furthermore propose a Wavelet Convolutional Block (WCB) to enhance the edge information extracting ca-pabilities, and also a Partial Channel Recalibration (PCR) block to allow mutual feature exchange between the two streams. We present experimental results for both methods on seven public datasets. Unlike conventional baselines that demonstrate convincing performance in some datasets and poor performance in others, the proposed segmenta-tion frameworks are able to consistently achieve state-of-the-art results on all datasets with very satisfactory F1 and IoU scores.