Bhattarai Binod ; Subedi Ronast ; Gaire Rebati Raman ; Vazquez Eduard ; Stoyanov Danail

Abstract

We present our novel deep multi-task learning method for medical imagesegmentation. Existing multi-task methods demand ground truth annotations forboth the primary and auxiliary tasks. Contrary to it, we propose to generatethe pseudo-labels of an auxiliary task in an unsupervised manner. To generatethe pseudo-labels, we leverage Histogram of Oriented Gradients (HOGs), one ofthe most widely used and powerful hand-crafted features for detection. Togetherwith the ground truth semantic segmentation masks for the primary task andpseudo-labels for the auxiliary task, we learn the parameters of the deepnetwork to minimise the loss of both the primary task and the auxiliary taskjointly. We employed our method on two powerful and widely used semanticsegmentation networks: UNet and U2Net to train in a multi-task setup. Tovalidate our hypothesis, we performed experiments on two different medicalimage segmentation data sets. From the extensive quantitative and qualitativeresults, we observe that our method consistently improves the performancecompared to the counter-part method. Moreover, our method is the winner ofFetReg Endovis Sub-challenge on Semantic Segmentation organised in conjunctionwith MICCAI 2021.