Archived

This content is available here for research, reference, and/or recordkeeping.

Author ORCID Identifier

https://orcid.org/0000-0003-0669-1407

Date Available

8-5-2026

Year of Publication

2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Engineering

Department/School/Program

Computer Science

Faculty

Abdullah-Al-Zubaer Imran

Faculty

Simone Silvestri

Abstract

Image segmentation is a fundamental task in computer vision. While segmentation models have traditionally been trained in a fully-supervised manner, recent approaches have leveraged large-scale pre-training and prompting mechanisms to great effect. However, the performance of such approaches often degrades when applied to domain-specific tasks like medical image analysis. A major reason for this lies is that these models are trained on large, labeled datasets of natural images, which have drastically different characteristics compared to medical images, limiting the generalizability of the methods when applied to medical data. This dissertation presents several data-efficient, adaptive, and promptable medical image segmentation models. Specifically, we introduce (1) methods that utilize promptable segmentation to refine coarse masks generated by models trained on extremely limited labeled data; (2) semi-supervised, multi-task learning methods that achieve strong performance on as few as five labeled samples and are capable of refining their own predictions; (3) adaptive probabilistic methods that capture the real-world variability of human annotators and incorporate multimodal information in the form of radiomics data; and (4) a multimodal, promptable method for segmentation of multiple organs from computed tomography scout images that carries real-world clinical implications.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.401

Archival?

Archival

Download

Share

COinS