Multi-Task Deep Learning for Image Segmentation Using Recursive Approximation Tasks

Fully supervised deep neural networks for segmentation usually require a massive amount of pixel-level labels which are manually expensive to create. In this work, we develop multi-task learning method relax constraint. We regard the problem as sequence approximation subproblems that recursively defined and in increasing levels accuracy. The handled by framework consists 1) task learns from ground truth masks small fraction images, 2) recursive conducts partial object regions data-driven mask evolution starting each instance, 3) other oriented auxiliary tasks trained with sparse annotations promote dedicated features. Most training images only labeled (rough) masks, do not contain exact boundaries, rather than their full masks. During phase, statistics these increased towards boundaries aided learned information fully fashion. network is on an extremely precisely segmented large set coarse labels. Annotations can thus be obtained cheap way. demonstrate efficiency our approach three applications microscopy ultrasound images.