Semantic Segmentation of Cardiac Structures from USG Images Using Few-Shot Prototype Learner Guided Deep Networks

This article proposed a method for semantic segmentation of ultrasound cardiac images using Deep network guided by learned prototype with few-shot learning approach. The main aim of the proposed method is to reduce the requirement for bulk training images to learn a supervised Deep neural network for semantic segmentation of cardiac structure. The proposed framework consists of two components, namely, a prototype learner and a supervised segmentor. The prototype learner generates a prototype for each type (class) of the cardiac structure, and the segmentor uses a U-Net architecture for segmentation and labeling of such structures. The prototype learner updates the parameters of feature extractor module by classifying the query set using K-nearest neighbor classifier and the support set. The output of the prototype learner (i.e., prototype) provides prior information about the classes. This output is fused with the output of the encoder of U-Net with an expectation that such a fusion will catalyze the learning process of the U-Net segmentor. Through training of the segmentor, the parameters of the encoding layer of U-Net get fine-tuned in a way so that it gradually aligns with the annotated segmentation maps. Probabilistic fusion model is used to amalgamate the prototypes with the generated features obtained from the encoder layer of the U-Net. To show the effectiveness of the proposed method, experiment was carried out on CAMUS dataset of 450 patients and performance is compared with those of four state-of-the-art techniques. To evaluate the performance of each of the methods, DICE index of endocardium, epicardium and atrium have been considered. Experimentation shows promising results obtained from the proposed technique.