Abstract

Medical imaging is routinely performed in clinics worldwide for the diagnosis and treatment of numerous medical conditions in children and adults. With the advent of these medical imaging modalities, radiologists can visualize both the structure of the body as well as the tissues within the body. However, analyzing these high-dimensional (2D/3D/4D) images demands a significant amount of time and effort from radiologists. Hence, there is an ever-growing need for medical image computing tools to extract relevant information from the image data to help radiologists perform efficiently. Image analysis based on machine learning has pivotal potential to improve the entire medical imaging pipeline, providing support for clinical decision-making and computer-aided diagnosis. To be effective in addressing challenging image analysis tasks such as classification, detection, registration, and segmentation, specifically for medical imaging applications, deep learning approaches have shown significant improvement in performance. While deep learning has shown its potential in a variety of medical image analysis problems including segmentation, motion estimation, etc., generalizability is still an unsolved problem and many of these successes are achieved at the cost of a large pool of datasets. For most practical applications, getting access to a copious dataset can be very difficult, often impossible. Annotation is tedious and time-consuming. This cost is further amplified when annotation must be done by a clinical expert in medical imaging applications. Additionally, the applications of deep learning in the real-world clinical setting are still limited due to the lack of reliability caused by the limited prediction capabilities of some deep learning models. Moreover, while using a CNN in an automated image analysis pipeline, it’s critical to understand which segmentation results are problematic and require further manual examination. To this extent, the estimation of uncertainty calibration in a semi-supervised setting for medical image segmentation is still rarely reported. This thesis focuses on developing and evaluating optimized machine learning models for a variety of medical imaging applications, ranging from fully-supervised, single-task learning to semi-supervised, multi-task learning that makes efficient use of annotated training data. The contributions of this dissertation are as follows: (1) developing a fully-supervised, single-task transfer learning for the surgical instrument segmentation from laparoscopic images; and (2) utilizing supervised, single-task, transfer learning for segmenting and digitally removing the surgical instruments from endoscopic/laparoscopic videos to allow the visualization of the anatomy being obscured by the tool. The tool removal algorithms use a tool segmentation mask and either instrument-free reference frames or previous instrument-containing frames to fill in (inpaint) the instrument segmentation mask; (3) developing fully-supervised, single-task learning via efficient weight pruning and learned group convolution for accurate left ventricle (LV), right ventricle (RV) blood pool and myocardium localization and segmentation from 4D cine cardiac MR images; (4) demonstrating the use of our fully-supervised memory-efficient model to generate dynamic patient-specific right ventricle (RV) models from cine cardiac MRI dataset via an unsupervised learning-based deformable registration field; and (5) integrating a Monte Carlo dropout into our fully-supervised memory-efficient model with inherent uncertainty estimation, with the overall goal to estimate the uncertainty associated with the obtained segmentation and error, as a means to flag regions that feature less than optimal segmentation results; (6) developing semi-supervised, single-task learning via self-training (through meta pseudo-labeling) in concert with a Teacher network that instructs the Student network by generating pseudo-labels given unlabeled input data; (7) proposing largely-unsupervised, multi-task learning to demonstrate the power of a simple combination of a disentanglement block, variational autoencoder (VAE), generative adversarial network (GAN), and a conditioning layer-based reconstructor for performing two of the foremost critical tasks in medical imaging — segmentation of cardiac structures and reconstruction of the cine cardiac MR images; (8) demonstrating the use of 3D semi-supervised, multi-task learning for jointly learning multiple tasks in a single backbone module – uncertainty estimation, geometric shape generation, and cardiac anatomical structure segmentation of the left atrial cavity from 3D Gadolinium-enhanced magnetic resonance (GE-MR) images. This dissertation summarizes the impact of the contributions of our work in terms of demonstrating the adaptation and use of deep learning architectures featuring different levels of supervision to build a variety of image segmentation tools and techniques that can be used across a wide spectrum of medical image computing applications centered on facilitating and promoting the wide-spread computer-integrated diagnosis and therapy data science.

Library of Congress Subject Headings

Image segmentation; Deep learning (Machine learning); Diagnostic imaging--Data processing

Publication Date

1-26-2023

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Cristian A. Linte

Advisor/Committee Member

Anthony Vodacek

Advisor/Committee Member

Mehdi Moradi

Campus

RIT – Main Campus

Plan Codes

IMGS-PHD

Download

Share

COinS