2026 Theses Doctoral

Robust Artificial Intelligence for Medical Image Analysis

Zhang, Xuzhe

Medical imaging is indispensable to healthcare, but the growing heterogeneities and developments of scanners, protocols, and techniques expose AI systems to an evolving data distribution, degrading the performance of the model trained on previous data distributions and hindering extensive deployments of AI in medical image analysis. This thesis focuses on addressing this technical gap with self- and semi-supervised frameworks that learn from heterogeneous and new data without the need for new manual annotations, enabling segmentation and quantification of image-based biomarkers that remain robust across study centers, imaging sequences, and protocols, even different generations of imaging modalities. The overarching and motivating goal is to develop robust AI model that could be applied to real-world, heterogeneous, and continually evolving data distributions.

We first proposed a unified domain adaptive segmentation framework named MAPSeg and conducted the first systematic evaluation of an unsupervised domain adaptation framework under different domain shifts and data access scenarios. MAPSeg is designed for practical deployment: it supports centralized training when data can be collected, privacy‑preserving federated learning when data cannot move, and test‑time adaptation when only unlabeled target data are available. It demonstrates superior robustness against common domain shifts in the neonatal infant brain segmentation task.

Building on that foundation, we built the first-ever fully automatic segmentation pipeline to segment pulmonary airway trees from MRI - an impossible task until the recent advancement of MRI acquisition sequence and reconstruction algorithm, and without using any annotation on MRI scans. The outcome framework enables radiation-free quantification of pulmonary airways that could pave the way toward repeated scans and longitudinal monitoring of the mismatch between airway caliber and lung size in the pediatric population, which was believed to arise early in life and is a risk factor of Chronic Obstructive Pulmonary Disease (COPD) in later life.

Finally, the work introduces Multimodal-DLEmph, a multimodal framework for standardized emphysema quantification on chest CT. Multimodal-DLEmph fuses image features with protocol‑aware priors and scanner metadata, and trains with distribution augmentations to learn protocol‑invariant representations, which results in a harmonized emphysema assessment that tracks patient trajectories consistently across centers and years, strengthening retrospective analyses, enabling fair pooling of heterogeneous cohorts, and improving the interpretability of imaging endpoints in observational studies and clinical trials.

Together, these contributions provide a principled and scalable path toward robust and trustworthy AI-assisted medical image analysis under domain shift. By unifying unsupervised adaptation, scalable self- and semi-supervised learning, as well as multimodal learning, the thesis advances AI-assisted computational tools that can broaden the reach of neurodevelopmental research, enable longitudinal quantitative airway studies from early life, and deliver reproducible biomarkers for chronic respiratory disease—ultimately helping translate AI from benchmarks to medical research and routine clinical practice.