Back to search

Cornell University

Making AI Accessible to Radiology: Robust Pre-processing, Segmentation, and Adaptation for Medical Imaging

Abstract

dc:description.abstract

This dissertation studies a question that limits the clinical usefulness of imaging AI: how can a model remain usable when data are heterogeneous, target structures are irregular, and deployment conditions change after training? I argue that accessibility in radiology AI is not achieved by a single high-performing model. Rather, it emerges when the pipeline, from input preparation, to segmentation representation, to post-deployment adaptation, to quantitative reporting, is designed to tolerate realistic variability without imposing excessive operational burden on clinicians. The thesis develops this argument through three methodological contributions and two application-oriented chapters. First, I present Neural Pre-Processing (NPP), a weakly supervised framework for end-to-end brain MRI pre-processing. NPP jointly addresses skull stripping, intensity normalization, and affine spatial normalization while separating geometry-preserving appearance correction from spatial transformation. This formulation improves speed and accuracy relative to traditional stepwise pipelines and makes downstream model inputs more usable. Second, I introduce InstaBound, a shape-agnostic approach to instance segmentation based on regression of an instance-aware signed distance field. Unlike methods that assume star-convexity or require multi-branch gradient heads, InstaBound uses a single branch to encode semantic and geometric information. This representation improves separation of touching structures while keeping the architecture simple. Third, I develop Seg-NuSA, a replay-free continual adaptation strategy for medical image segmentation. Seg-NuSA constrains updates to low-interference subspaces derived from pretrained weights, reducing catastrophic forgetting under privacy-constrained deployment. Across cross-modality and sequential adaptation settings, it yields a stronger retention-adaptation trade-off than fine-tuning and regularization-based baselines. These methodological chapters are then connected to a translational program in autosomal dominant polycystic kidney disease (ADPKD). I synthesize work on test-retest reproducibility, MRI quality control, liver cyst segmentation, pancreatic cyst detection, and the TraceOrg platform for automated measurement of kidney, liver, and cyst volumes. I then show how these workflow advances enable biomarker studies including growth-rate modeling, water therapy, tolvaptan response assessment, pregnancy-associated cyst dynamics, and spleen contouring in myelofibrosis. Taken together, the dissertation advances a framework for accessible radiology AI in which technical robustness, deployment stability, and measurement reproducibility are treated as mutually dependent design goals rather than independent stages of a pipeline.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Biomedical Engineering
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Biomedical Engineering
Grantor
Cornell University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • He, Xinzi
Chair dc:contributor.chair
  • Sabuncu, Mert
Committee members dc:contributor.committeemember
  • Doerschuk, Peter
  • Zabih, Ramin

Subjects

dc:subject × 4

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Other
He_cornellgrad_0058F_15572
https://www.proquest.com/LegacyDocView/DISSNUM/32676319

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

He, Xinzi. Making AI Accessible to Radiology: Robust Pre-processing, Segmentation, and Adaptation for Medical Imaging. Doctor of Philosophy thesis, Cornell University, 2026. https://hdl.handle.net/1813/126533