U-Net Based Deep Learning Approach for 2D Cardiovascular Image Segmentation

The goal of medical image segmentation is to organize pixels into several areas according to the several characteristics of the images. Due to several factors, such as variations in data signal-to-noise ratios, signal intensities, and individual variations in heart morphologies, segmenting 2D echo cardiovascular images remains a difficult process. This research introduces 3D U-Net-SparseVoxNet, a unique and effective 3D sparse convolutional network based on U-Net. Any two layers in this network that have the same feature map size may have direct connections with each other, and there are fewer connections overall. Consequently, by drastically reducing the network depth and ultimately utilizing a spatial selfattention mechanism to improve feature representation, 3D U-Net-SparseVoxNet can successfully handle the optimization issue of gradients disappearing when using a limited sample of data to train a 3D deep neural network architecture. This research presents a detailed evaluation of the suggested technique using the HVSMR 2016 dataset. The strategy performs better when compared to other approaches. The proposed method proved to provide good and efficient results in classifying the data with an accuracy of 90% compared to 3D U-Net and VoxResNet, with 74% and 80% accuracy, respectively.