Selective Compression of Medical Images via Intelligent Segmentation and 3D-SPIHT Coding

Abstract

ABSTRACT SELECTIVE COMPRESSION OF MEDICAL IMAGES VIA INTELLIGENT SEGMENTATION AND 3D-SPIHT CODING by Bohan Fan The University of Wisconsin-Milwaukee, 2018 Under the Supervision of Professor Zeyun Yu With increasingly high resolutions of 3D volumetric medical images being widely used in clinical patient treatments, efficient image compression techniques have become in great demand due to the cost in storage and time for transmission. While various algorithms are available, the conflicts between high compression rate and the downgraded quality of the images can partially be harmonized by using the region of interest (ROI) coding technique. Instead of compressing the entire image, we can segment the image by critical diagnosis zone (the ROI zone) and background zone, and apply lossless compression or low compression rate to the former and high compression rate to the latter, without losing much clinically important information. In this thesis, we explore a medical image transmitting process that utilizes a deep learning network, called 3D-Unet to segment the region of interest area of volumetric images and 3D-SPIHT algorithm to encode the images for compression, which can be potentially used in medical data sharing scenario. In our experiments, we train a 3D-Unet on a dataset of spine images with their label ground truth, and use the trained model to extract the vertebral bodies of testing data. The segmented vertebral regions are dilated to generate the region of interest, which are subject to the 3D-SPIHT algorithm with low compress ratio while the rest of the image (background) is coded with high compress ratio to achieve an excellent balance of image quality in region of interest and high compression ratio elsewhere.