A U-Net with Statistical Shape Restrictions Applied to the Segmentation of the Left Ventricle in Echocardiographic Images

Abstract

This paper aims to introduce an innovative approach to semantic segmentation by leveraging a convolutional neural network (CNN) for predicting the shape and pose parameters of the left ventricle (LV). Our approach involves a modified U-Net architecture with a regression layer as the final stage, as opposed to the traditional classification layer. This modification allows us to predict all the shape and pose parameters of a statistical shape model, including rotation, translation, scale, and deformation. The adapted U-Net is trained using data from a point distribution model (PDM) of the LV. The experimental results demonstrate a mean Dice coefficient of 0.82 on good quality images, and 0.66 including mean and low-quality images. Our approach successfully overcomes a common issue encountered in CNN-based semantic segmentation. Unlike the inaccurate pixel classification that often leads to unwanted blobs, our CNN generates statistically valid shapes. These shapes hold significant potential in initializing other methods, such as active shape models (ASMs). Our novel CNN-based approach provides a novel solution for semantic segmentation, offering shapes and pose parameters that can enhance the accuracy and reliability of subsequent medical image analysis methods.