Date of Award
Summer 2021
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Ye, Dong Hyed
Second Advisor
Povinelli, Richard J.
Third Advisor
Medeiros, Henry
Abstract
In general radiography, inconsistency of brightness and contrast in initial presentation is a common complaint from radiologists. Inconsistencies, which may be a result of variations in patient positioning, dose, protocol selection andimplant could lead to additional workflow by technologists and radiologists to adjust the images. To tackle this challenge posed by conventional histogram-based display approach, first a Deep Learning based Brightness Contrast (DL BC) algorithm to improve the consistency in presentation by using a residual neural network trained to classify X-ray images based on a novel NxM grid of brightness and contrast combinations is proposed. More than 8,500 unique images from sites in US, Ireland and Sweden covering 27 anatomy/view combinations were used for training. The DL BC model achieved an average test accuracy of 99.2% on a set of 2700 images. Quantitative evaluation of DL BC using ROI based metrics on a set of 140 images covering 7 anatomies showed a 35.3% improvement in mean pixel intensity consistency and a 15.5% improvement in contrast consistency. Secondly, we propose a Deep Learning based Metal Segmentation (DL MS) approach using U-Net architecture that is capable of segmenting highly attenuating metal regions from an image. By this method, we were able to exclude highly attenuating regions from anatomy mask during conventional histogram-based display calculation and hence avoid potential inconsistencies. On comparing quantitative evaluation results of DL BC and DL MS on a set of 12 Wrist PA images, it was observed that DL BC is sensitive to image orientation and introduced 22.3% shift in contrast, whereas DL MS introduced just 10.3% shift in contrast. When compared with conventional approaches, both proposed DL BC and DL MS approach demonstrate the feasibility of using deep learning techniques to reduce inconsistency in initial display presentation and improve user workflow.