This chapter discusses the importance of machine learning in the field of medical imaging for reconstructing medical images from the measured raw data. Besides discussing the use of machine learning for medical image reconstruction, a general overview is also provided on all the existing techniques of medical imaging. Mathematical models are provided in order to understand better the use of machine learning for reconstruction purposes. We discuss both unsupervised techniques like dictionary learning, auto-encoders, and supervised techniques which include learning of hyperparameters and various regularization methods used in deep learning models that replace various steps in iterative algorithms used for image reconstruction such as for image enhancement.
Computerized tomography scan (or CT scan) is one of the most prominent diagnostic tools used in medical imaging which provides anatomical information about the human body. It uses X-ray machines that rotate and produce 2-dimensional images of the sections of the body under observation and can be visualized easily on the computing devices. There are various deep learning approaches used to reconstruct the medical images by the use of neural networks. In this paper, deep learning based Generative Adversarial Networks (or GAN's) methods have been proposed to reconstruct images from a given randomized vector and trained data. The main goal of the proposed model is to train a better generator than a discriminator in such a way that the reconstructed images are as real as possibly present in the training data.
Convolutional neural networks (CNN) have become a popular choice for image segmentation and classification. Internal body images are obscure in nature with involvement of noise, luminance variation, rotation and blur. Thus optimal choice of features for machine learning model to classify bleeding is still an open problem. CNN is efficient for attribute selection and ensemble learning makes a generalized robust system. Capsule endoscopy is a new technology which enables a gastroenterologist to visualize the entire digestive tract including small bowel to diagnose bleeding, ulcer and polyp. This paper presents a supervised learning ensemble to detect the bleeding in the images of Wireless Capsule Endoscopy. It accurately finds out the best possible combination of attributes required to classify bleeding symptoms in endoscopy images. A careful setting for CNN layer options and optimizer for back propagation after reducing the color palette using minimum variance quantization has shown promising results. Results of testing on public and real dataset has been analyzed. Proposed ensemble is able to achieve 0.95 on the public endoscopy dataset and 0.93 accuracy on the real video dataset. A detailed data analysis has also been incorporated in the study including RGB pixel intensities, distributions of binary classes and various class ratios for training.