Thesis

Semi-supervised medical image segmentation with ADMM-based energy regularization

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T17980
Person Identifier (Local)
  • 202470646
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • This paper addresses the challenge of data scarcity in medical image segmentation by introducing a suitable energy functional as a loss term within a semi-supervised learning framework. Inspired by the variational model proposed by Wei et al. (2017), in which the Potts energy functional is minimized using a region force derived from the negative log-likelihood of the Bernoulli distribution, this study explores the integration of such energy formulations with a U-Net–based deep learning model. Unlike the classical Chan–Vese model, which is prone to local minima due to its non-convex nature, the convex relaxed Potts energy model enables global optimization. However, when directly employed as a loss function, the energy functional remains highly nonlinear and leads to suboptimal convergence. To overcome this limitation, an ADMM-based optimization strategy, following the approach of Wei et al., is adopted to enhance convergence stability and computational efficiency. When incorporated as a regularization term in the neural network, the proposed energy-based loss serves as a spatial prior that mitigates the limitations of deep learning in enforcing spatial coherence and boundary consistency, particularly for irregular or noisy lesion boundaries. Moreover, this integration effectively reduces the dependence on large-scale labeled data, thereby improving the applicability of deep models to semi-supervised segmentation scenarios. Comprehensive numerical experiments demonstrate that the proposed ADMM-enhanced semi-supervised method achieves accurate and stable segmentation performance under limited annotation conditions.
Advisor / supervisor
  • Chen, Ke, 1962-
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17980 2026-04-27 Public Download