U-Net adaptation for multiple instance learning

Abstract

Multiple instance learning (MIL) is a weakly supervised learning method where a single label is assigned to a group of instances. Recent advancement in neural networks makes it possible to achieve great results but the training requires many annotated examples which can be difficult to obtain. In case of medical imaging, such a method can theoretically provide voxel-level annotations basing on the image-level annotations. More precisely, taking a training dataset where each image is given with a label indicating a presence of a region of interest (ROI), the model can be trained to produce voxel labels of the object. We propose a modification of the U-Net architecture for image segmentation that can be trained on weakly labeled datasets and solves the MIL problem. The U-Net architecture is famous for being able to train on small number of samples which can further improve benefits of MIL approach. In this paper we also present results of experiments on synthetic data with investigation of stability to noise and medical images. Experimental results prove that suggested modifications improve the noise stability of the weakly supervised method.