Date of Award

January 2021

Document Type


Degree Name

Medical Doctor (MD)



First Advisor

Julius Chapiro


Hepatocellular Carcinoma (HCC) is the most common primary liver cancer worldwide and the fastest rising cause of cancer deaths in the United States [1]. HCC frequently arises within the context of pre-existing liver conditions such as Hepatitis B or C virus (HBV or HCV) infection, chronic alcohol abuse, or nonalcoholic fatty liver disease (NAFLD). NAFLD is increasingly important in the pathogenesis of HCC, as it is the fastest rising cause of cancer deaths within the United States, with incidence increasing approximately 3-fold since the 1980’s [2]. Image guided therapies are important for the treatment of HCC, particularly for patients who are not candidates for surgical resection or liver transplantation [3]. Thermal ablative techniques such as microwave ablation (MWA) and radiofrequency ablation (RFA) can induce coagulation of tumor cells through the direct application of energy to the tumor, and are frequently performed under US or CT guidance. MWA and RFA can serve as curative therapy for early-stage HCC. Transarterial chemoembolization (TACE) is another popular treatment for HCC which involves administration of chemotherapy intra-arterially directly to a tumor site, followed by embolization of said tumor. TACE is performed under fluoroscopic guidance, and is the most commonly used treatment for intermediate-stage HCC [4]. Other locoregional treatment options for HCC include Y-90 Radioembolization, combination TACE + RFA/MWA, as well as TACE + sorafenib or immune checkpoint inhibitors. Unfortunately, the rate of distant recurrence with HCC after ablative and intra-arterial therapies remains high. Distant recurrence, or the presence of a new intra-hepatic nodule or recurrence of HCC outside of the liver after an ablation, is common. Studies have shown that the rate of remote intrahepatic or extrahepatic metastasis after RFA can range from 42-60% [5]. Data for MWA is scarcer, but distant recurrence after this treatment modality in certain populations has been shown to be around 40% [5]. The risk of local recurrence, or recurrence at a location peripheral to the ablation cavity, is lower than the risk of distant recurrence. With both RFA and MWA, the recurrence rate is similar at around 10% [5]. For TACE, the chance of recurrence after initial treatment is also a concern. Studies have reported recurrence rate after TACE to be as high as 42% [6]. Several risk factors have been reported to correlate with HCC recurrence after locoregional therapy, including but not limited to size and number of tumors, AFP levels, and Child-Turcotte-Pugh scores [7]. Locoregional therapies themselves might also change the tumor microenvironment and result in pro-oncogenic reactions which influence tumor recurrence both locally and distally [8-10]. The risk of HCC recurrence after ablation remains a clinically relevant question as it can influence transplant organ allocation, optimal imaging follow-up strategies, and choice of adjuvant therapy. A recent surge in the availability of low-cost computational power has made a machine learning approach to image analysis, as well as computer-aided diagnosis and prognosis broadly accessible [11]. This offers a promising new avenue for the exploration of risk quantification after interventional procedures for HCC. Patients who undergo locoregional therapy for HCC have imaging studies performed both before and after any intervention is performed, generating large volumes of data. These data sets can be analyzed using a variety of machine learning techniques to generate imaging features which would otherwise be inaccessible to a human observer. These data sets can be trained using various machine learning techniques to produce models which offer predictive insight into the risk of HCC recurrence. There are three overarching goals of this thesis. The first is to provide a background for HCC epidemiology and treatment, specifically with respect to early- and intermediate-stage disease. Discussions on HCC treatment will focus primarily on minimally invasive image-guided therapy, as a detailed discussion on surgical approaches are outside the scope of this thesis. This discussion will help contextualize the importance of additional research findings described later. The second goal of this thesis is to demonstrate that combination TACE+TA treatment is not necessary for HCC smaller than 3cm. Combination TACE+TA treatment has been shown to be efficacious for larger HCCs, but its role in the treatment of smaller HCC has not been fully elucidated. The third goal of this thesis is to establish a proof-of-concept demonstration that machine learning techniques can be used to offer predictive insight into HCC recurrence after locoregional therapy. This model will be contextualized within the broader landscape of existing machine learning approaches to medical imaging diagnosis of HCC. This will demonstrate that machine learning is important not just within the realm of diagnostic radiology, but will also serve a key role within interventional oncology.


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