Ph.D. Dissertation Defense: S M Kamrul Hasan
Ph.D. Dissertation Defense
From Fully-Supervised Single-Task to Semi-Supervised Multi-Task Deep Learning Architectures for Segmentation in Medical Imaging Applications
S M Kamrul Hasan
Imaging Science Ph.D. Candidate
Chester F. Carlson Center for Imaging Science, RIT
While deep learning-based image analysis has the potential to significantly improve the medical image acquisition to diagnosis or therapy pipeline, many of these successes are achieved at the cost of a large pool of labeled datasets, which requires substantial domain expertise and manual labor. Furthermore, the applications of deep learning in the clinical settings are still limited due to inadequate reliability of the deep learning-based model predictions; one example is the generation of sufficiently accurate segmentation maps. This dissertation proposes the development of optimized and annotation-efficient deep learning models for a variety of medical imaging applications, including multi-modality image segmentation, uncertainty estimation, and image registration. We proposed a novel tool for segmenting and digitally inpainting surgical instruments from laparoscopic video sequences routinely used during robot-assisted interventions using fully-supervised, single-task learning. Similarly, we also implemented an unsupervised learning-based deformable registration framework for constructing patient-specific cardiac models from cine cardiac magnetic resonance (MRI) images using segmentation masks generated using our novel memory-efficient algorithm that leverages learned group convolution and weight pruning techniques. To estimate the uncertainty associated with the obtained segmentation masks in the effort to inform the clinicians whether or where the segmentation masks need correction, we incorporate a Monte Carlo dropout into our fully-supervised model to transform it into a Bayesian version. We further extended our work into a semi-supervised, single-task learning-based approach to deal with the pseudo-label bias problem by leveraging self-training (through meta pseudo-labeling) in concert with a Teacher network that instructs a Student network by generating pseudo-labels given unlabeled input data. We also developed a largely unsupervised, multi-task learning framework to demonstrate that disentanglement together with mutual information minimization can improve segmentation performance even with 1% of the training data, while combined with a generative adversarial network and a conditioning layer-based reconstruction network. Lastly, to further improve model generalizability, we proposed a 3D semi-supervised, multi-task learning framework for jointly learning multiple tasks in a single backbone module – uncertainty estimation, geometric shape generation, and feature segmentation – and demonstrated its performance in segmenting the left atrial endocardial surface from 3D MRI images. In summary, this dissertation demonstrates the implementation, integration and adaptation of deep learning architectures featuring different levels of supervision to build a variety of image segmentation 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.
Undergraduates, graduates, and experts. Those with interest in the topic.
To request an interpreter, please visit https://myaccess.rit.edu
When and Where
Open to the Public