Date of Award

January 2020

Document Type


Degree Name

Medical Doctor (MD)



First Advisor

Julius Chapiro, MD

Second Advisor

David Madoff, MD


The aim of this two part study was to develop automated methods of analyzing clinical imaging to predict treatment response to trans-arterial chemoembolization (TACE).

Part I

Part I aimed to develop an automated method to quantify the volume of Lipiodol within a target tumor after conventional trans-arterial chemoembolization (cTACE) and evaluate this volume as a potential early marker of treatment response. A retrospective cohort of 63 patients with metastatic liver tumors of neuroendocrine origin treated with cTACE was analyzed. Volumetric masks of total Lipiodol (ethiodized oil) coverage were generated from post-embolization non-contrast Computed Tomography (CT) by an algorithm that classified tumor voxels as Lipiodol if they exceeded the reported upper limit of normal liver parenchyma, 67 Hounsfield Units (HU). “High density” Lipiodol deposits were similarly measured by applying a higher HU threshold (200 HU). By overlaying 3D masks, the volumes of Lipiodol (total and high density) within tumor tissue (total and enhancing) were compared after cTACE between treatment response groups as determined by quantitative European Association for the Study of Liver (qEASL) radiographic response criteria. The percent of target tumor volume and enhancing tumor volume covered by Lipiodol was higher in cTACE responders compared to non-responders (p=0.008, p= 0.019 respectively). The percent of enhancing tumor tissue covered by high density Lipiodol also differed between responders and non-responders (47.9% vs. 4.5%; p<0.001) with the largest effect size of all comparisons made (Cohen’s term d=2.02). A Receiver Operating Characteristic (ROC) curve identified 12.57% of enhancing tumor tissue covered by high density Lipiodol as the cutoff best differentiating treatment responders from non-responders (AUC=0.77, sensitivity 76.92%, specificity 83.33%). An automated method of quantifying volumetric Lipiodol within a target tumor may serve an early imaging biomarker of treatment response to cTACE, quantifiable within 24 hours of treatment.

Part II

Part II aimed to use magnetic resonance (MR) imaging and clinical patient data to predict therapeutic outcomes of TACE by applying machine learning (ML) techniques. This study included 36 patients with hepatocellular carcinoma (HCC) treated with TACE. Image-based tumor response to TACE was assessed using 3D quantitative response criteria (quantitative European Association for the Study of the Liver criteria; qEASL). Clinical information, baseline imaging, and therapeutic features were used to train logistic regression (LR) and random forest (RF) models to predict patients as treatment responders or non-responders. The performance of the models was compared to treatment response as determined by qEASL response criteria using leave-one-out cross-validation. Both LR and RF learning models achieved an overall accuracy of 78% (sensitivity 62.5%, specificity 82.1%, PPV 50.0%, NPV 88.5% for both) when predicting TACE treatment response. The strongest predictors of treatment response observed (78.8% accuracy vs. 72% accuracy in the next best model) included a clinical variable (presence of cirrhosis) and imaging variable (tumor signal intensity >27.0). TACE outcomes in patients with HCC may be predicted pre-procedurally by combining clinical patient data with baseline MR imaging using ML techniques.


This thesis is restricted to Yale network users only. This thesis is permanently embargoed from public release.