Predicting Survival Prior to Surgery in Colorectal Liver Metastases Using Radiomics

Abstract

Colorectal cancer has the second highest mortality rate of all cancers globally, with the development of distant metastases heavily contributing to this. The most common site of metastasis from colorectal cancer is the liver. Although patients may receive curative intent resection of their colorectal liver metastases, many patients will experience recurrence. Additional therapies may be given to higher risk patients, but determining which patients are high risk compared to others remains a difficult clinical problem. Reliable prognostic models of death or recurrence following hepatectomy of colorectal liver metastases are critical for creating individualized treatment plans. Clinical models have been created for this purpose in the past, but have shown limited generalizability. Radiomic models can take advantage of radiographic imaging that is routinely collected for this disease, but radiomic biomarkers may be affected by the protocol used to capture the images. The correction of these influences may affect downstream model performance. As both clinical and radiomic features have demonstrated their potential in prognostic modeling for colorectal liver metastases, clinicoradiomic prognostic models have been proposed. However, these studies often lack consideration for benchmarking and the potential differences in performance feature correction can cause. In this thesis, we developed several types of prognostic models for overall and hepatic disease-free survival using hand-crafted and deep learning radiomic features, clinical features, and a mixture thereof. We also provided insights into the use of ComBat for radiomic feature correction and its effects on model performance. All models were developed using a large, multi-institutional cohort of 1,301 patients. Through our analysis, we have shown that models which integrate radiomic features can prognosticate outcome and that clinicoradiomic models can outperform clinical models for both survival outcomes. We have also shown that radiomic features are affected by imaging slice thickness and that when these features are corrected using ComBat, they generally produce lower performing models. This work demonstrates the utility of radiomic features for the prognostication of overall and hepatic disease-free survival and how radiomics can compliment clinical information to improve risk prediction and treatment decision-making for patients with colorectal liver metastases.