Design of U-Net Architectures for Medical Image Segmentation using AI/ML Model

Abstract

The basic digital healthcare is medical image segmentation. Required to diagnose disease with computers and for computer-assisted surgical planning to monitor treatment over the years. Fully convolutional encoder-decoder networks (FCN) have emerged as the standout framework in dense pixel-level classification among the deep learning techniques. Fully convolutional encoder-decoder networks, in particular U-Net, are the leading method for dense pixel-level classification. The distinguishing features of U-Net are its symmetrical contracting and expanding pathways, with skip connections via high-resolution dense intermediaries. That layout matters. It provides for maintaining precise localization of the network and high segmentation accuracy even in the case of a limited amount of samples in annotated datasets and their data is limited. Such flexibility is crucial in biomedical imaging where it is difficult to overcome the challenges of large variations between patients in the properties of tissues and their spatial position. It also involves the changes in shape, volume, anatomical orientation and weak natural contrast at the boundaries. The idea has matured over the past decade to five distinct architectural configurations: standard U-Nets, dedicated deep nets, structural variations, hybrid models (hUNet) and ensembles (eUNet). These variants put in different custom feature extract encoders as well as better local decoding blocks. Some also further optimize the loss functions to improve the semantic segmentation accuracy. Some of these variants have already been applied to clinical applications such as to the precise cardiac volume profiling of CT images using U-shaped Generative Adversarial Networks (GANs), deep urinary system analysis and localization of calculus in abdominal radiography by a computer. Data is seldom abundant. This ongoing training choke is alleviated with newer data augmentation techniques which are incorporated into the pipeline on a regular basis. In addition to these progressions in artificial intelligence modeling, parameter optimization and structured network pruning, an increasing number of simulations are now being executed at speed. The models consume less VRAM, require less storage when used in clinical workflow, infer faster, and they do not have any black-box problem and that makes the faster and more accurate in segmentation.