CNN Assisted Hybrid Algorithm for Medical Images Segmentation

In this report we focus on a hybrid method based on a convolutional neural network (CNN) for histological image segmentation. We propose a CNN assisted interactive segmentation tool with weakly-supervised learning to accelerate the process of manual image annotation. The core of our annotation approach is a classical KNN classifier that uses parameters predicted by CNN. User annotates an image with scribbles of two types corresponding to glands and non-glands histological structures. Next the model performs label propagation to all unlabeled pixels providing user a fully annotated image build from his scribbled-based input. User can interact with the annotation tool and add new scribbles to correct the result. The algorithm allows to reduce one image annotation time from 150 to 25-30 minutes for PATH-DT-MSU dataset that can potentially seriously increase the number of fully annotated histological images which is necessary for the development of diagnostic algorithms.