Thesis
Supporting the infrastructure for developing and deploying AI/ML models In clinical systems
University of Iowa
Master of Science (MS), University of Iowa
Spring 2026
DOI: 10.25820/etd.008373
Abstract
Machine learning in clinical imaging has demonstrated substantial promise in stroke care, where rapid and accurate decision-making is critical. Machine and deep learning systems can reduce time-to-diagnosis, alleviate documentation burden, and provide consistent analytical support for clinical decisions. However, despite strong performance in research settings, few systems are translated into routine clinical use. A primary challenge for deploying into clinical use is due to limitations of research focused development pipelines, including limitations in foundational imaging tooling, constraints in realistic model development workflows, and a lack of deployment mechanisms that use well-established pathways. Overcoming these barriers requires coordinated improvements across each stage of development. This thesis addresses these bottlenecks by strengthening foundational tooling, extending multimodal image model development workflows, and developing and validating a Fast Healthcare Interoperability Resources (FHIR)-compliant integration pathway for clinical deployment.
At the foundational level, ITK serves as a core open-source library for medical image analysis. However, its testing and wrapping infrastructure for native C++ modules into Python relied on outdated documentation and convoluted workflows, limiting extensibility and creating barriers for developers. To address this, I modernized and formally documented the processes for uploading testing data and building custom modules within ITK’s existing framework. These improvements strengthened maintainability and extensibility, reinforcing ITK as a foundation for clinical imaging research and system development.
At the model development stage, frameworks such as MONAI Label support interactive annotation and model training, reducing the time and effort required to generate labeled datasets. However, the framework did not natively support coordinated management of multi-modal datasets, which limited its use for models that rely on complementary modalities. An extension to the MONAI Label was implemented to enable simultaneous annotation of complementary imaging volumes within a unified workflow. This enhancement enables multimodal annotation and training within the existing framework, supporting the development of clinically relevant multimodal imaging models.
At the deployment stage, clinical imaging systems must integrate with existing electronic health record (EHR) and Picture Archiving and Communication Systems (PACS) to support routine clinical use. However, developing and validating such integrations is challenging because commercial EHR and PACS systems are proprietary and difficult to access for research and testing. To enable feasible development and evaluation, a FHIR-compliant module within the analysis pipeline was integrated with Medplum, an open-source EHR platform supporting both local development and enterprise deployment. This approach enabled standardized FHIR resources to be generated, transmitted, and retrieved within a local environment, demonstrating a pathway for clinical integration in production environments.
By strengthening foundational tooling extensibility, enabling multimodal model development, and validating the feasibility of a process for system integration within EHR/PACS environments, this work reduces fragmentation across the clinical imaging AI pipeline. Together, these contributions provide a more continuous and reproducible pathway for developing, validating, and integrating clinical imaging tools.
Details
- Title: Subtitle
- Supporting the infrastructure for developing and deploying AI/ML models In clinical systems
- Creators
- Cavan Riley
- Contributors
- Gary Christensen (Advisor)Hans Johnson (Committee Member)Tyler Bell (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Spring 2026
- DOI
- 10.25820/etd.008373
- Publisher
- University of Iowa
- Number of pages
- xi, 77 pages
- Copyright
- Copyright 2026 Cavan Riley
- Language
- English
- Date submitted
- 04/23/2026
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 69-77).
- Public Abstract (ETD)
Machine learning shows strong potential in clinical imaging, yet few systems reach routine hospital use due to fragmented development pipelines. This work addresses three bottlenecks across that pipeline. First, the Insight Toolkit’s (ITK) testing and Python packaging infrastructure was modernized to improve extensibility for downstream developers. Second, MONAI Label was extended to support multimodal imaging datasets, enabling coordinated annotation and model training across complementary MRI volumes. Third, a FHIR-compliant integration pathway was developed using open-source tools, demonstrating automated delivery of analysis results into electronic health record and imaging systems. Together, these contributions create a more complete pathway from model development to clinical deployment, increasing the likelihood of system adoption.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9985177375402771
Metrics
1 Record Views