Leveraging Radiomics Across the Oncological Journey: From Risk Stratification to Metastasis Prediction

David Martinus (Advisor: Eugene Koay, PhD

Hepatocellular carcinoma (HCC) and pancreatic ductal adenocarcinoma (PDAC) are highly lethal gastrointestinal malignancies for which current guidelines rely predominantly on qualitative radiological assessments and standard clinical variables. Radiomics provides a quantitative framework to extract high-dimensional features from routine contrast-enhanced CT that have potential to capture subvisual tissue patterns and heterogeneity relevant to cancer initiation, progression, and metastatic spread. In this dissertation, we developed and evaluated radiomics pipelines spanning image preprocessing, automated segmentation, feature extraction, harmonization, and machine learning to address three key problems across the oncologic trajectory: HCC risk stratification in cirrhosis, characterization of pancreatic cystic lesions, and prediction of liver metastasis and survival in PDAC.

Using multi-center cohorts, we first compared venous phase liver radiomics features to an established clinical risk index in patients with cirrhosis to risk stratify patient progression to HCC. We demonstrated that the combination of radiomics features with the clinical risk model improved discrimination, time-dependent AUC, and yielded high C-index in external validation. The improvement in performance of the combined model over the clinical risk index alone indicates that liver radiomics encode information that can be used to enhance HCC risk stratification in patients with cirrhosis.

We then applied radiomics features from pancreas, pancreatic cyst, and body fat region segmentations to differentiate mucinous from non-mucinous pancreatic cysts and to estimate malignant potential in patients with intraductal papillary mucinous neoplasms (IPMNs). We compared our radiomics model to the clinical data and 'worrisome' imaging features that guide current clinical management of IPMNs. We quantified the impact of radiomics differences across multiple institutions despite data harmonization techniques and reported performance attenuation in blinded external validation cohorts.

Finally, we built radiomics and clinical models using pancreas and liver features to predict distant metastasis-free survival in PDAC patients treated with neoadjuvant chemotherapy or upfront surgery. We observed that combined radiomics-clinical models consistently achieved the highest or near-highest C-indices, significant Kaplan–Meier separation of risk groups, and metastasis-specific information beyond conventional variables alone.

Collectively, these studies demonstrate that radiomics features can synergize with clinical and laboratory data to enhance risk stratification, cyst classification, and metastasis prediction in HCC and PDAC. We outline the practical challenges and suggest techniques to improve generalizability of these models for future clinical translation.

Advisory Committee:

• Eugene Koay, PhD, Chair
• Kristy Brock, PhD
• Suprateek Kundu, PhD
• Radhe Mohan, PhD
• Surendra Prajapati, PhD
• Simona Shaitelman, MD

Join via Zoom (Please contact Mr. Martinus for his Zoom meeting info.)