Colorectal cancer liver metastases: Improving treatment strategies of patients using imaging-based artificial intelligence

The aim of this thesis was to improve treatment strategies of patients with colorectal cancer liver metastases (CRLM) using imaging-based artificial intelligence. This thesis presents studies on radiomics models predicting patient tumor characteristics and outcomes, on the inclusion of new imaging features in tumor response evaluation, and on automating tumor response assessment. In Chapter 2, an overview is provided of the literature on the value of radiomics in predicting response to treatment in patients with gastrointestinal tumors. In this comprehensive systematic review, a total of 60 studies are discussed. In Chapter 3, the current state and potential of radiomics to predict clinical outcomes, including response to treatment, recurrence of disease, and survival, in patients with CRLM is investigated. In Chapter 4, total tumor volume (TTV) response after systemic treatment is compared to tumor response according to RECIST1.1 in patients with initially unresectable CRLM of the randomized controlled CAIRO5 trial. In addition, the prognostic value of TTV response and RECIST1.1 for recurrence-free survival in a subgroup of patients with secondarily resectable CRLM after induction treatment is examined. In Chapter 5, the interobserver variability between multiple observers performing morphologic tumor response assessment according to set criteria is evaluated in patients with CRLM treated with systemic therapy. In Chapter 6, the development of automatic segmentation deep learning models for automatic TTV assessment of patients with CRLM is described. In Chapter 7, the prognostic value of TTV is further assessed for early recurrence (within 6 months) and overall survival in a subgroup of patients from the CAIRO5 cohort treated with induction systemic therapy followed by complete local treatment. The findings are validated in an external cohort of patients with CRLM treated with induction chemotherapy at Erasmus Medical Center Rotterdam. In Chapter 8, clinical patient characteristics, including pre-treatment TTV, are compared between patients with different KRAS mutation variants in a subgroup of the CAIRO5 cohort who carried a KRAS mutation and were treated with bevacizumab and chemotherapy. In Chapter 9, the relationship between radiomics imaging features and genetic mutation status (KRAS versus wildtype) is investigated in patients with CRLM from the CAIRO5 cohort. The findings are validated in an external cohort of patients from the National Cancer Institute. Finally, this thesis concludes with a general discussion and future perspectives in Chapter 10