Study Design and Patient Selection
This institutional review board-proved study was performed as a retrospective single-center observational trial in a university liver center. Informed consent was waived due to the retrospective character of the study. All patients with TARE for hepatic mCRC and available pre-TARE imaging as well as follow-up data for response evaluation (n = 51), identified from the entire cohort with TARE (n = 288), in our center between 2009 and 2021 were included. All patients had experienced stable or progressive disease under prior systemic therapy and were assigned for TARE after an interval of minimum 4 weeks after the end of systemic therapy. Pre-TARE imaging was defined as the last available contrast enhanced CT before TARE, in this case at least 4 weeks after the end of systemic therapy and a maximum of 4 weeks before TARE. Treatment response was assessed by the Response Evaluation Criteria in Solid Tumors (RECIST 1.1) at follow-up after 3 months. Patients were stratified into responder (complete/partial response and disease control) and non-responder (progressive disease).
Procedure Details
First, evaluation and preparation of TARE was performed by using Tc-99m-MAA following embolization of aberrant vasculature originating from the hepatic circulation. Subsequent to Tc-99m-MAA application, patients underwent planar whole body and SPECT/CT scanning of the thoracic and abdominal region (Discovery NM630, GE Healthcare, Solingen, Germany) using low-energy collimators for detection of extrahepatic tracer accumulation, and assessment of hepatopulmonary shunting. The maximum pulmonary shunt fraction accepted was 20%. In case of positive TARE evaluation, TARE using resin microspheres (SIR-Spheres®; Sirtex Medical, Sydney, Australia) was performed according to standard operating procedures.
Data Collection
Electronic patient records have been reviewed for patient characteristics, including age, sex, pre-therapeutic laboratory data, primary tumor side (left vs. right hemicolon), metastatic tumor burden of the liver, previous therapies, and KRAS status (wild-type or mutated). Primary tumors arising in the splenic flexure, descending colon, sigmoid colon, or rectum were classified as left-sided mCRC; tumors arising in the appendix, cecum, ascending colon, hepatic flexure, or transverse colon were classified as right-sided mCRC.
Image segmentation and feature extraction
Two experienced radiologists, blinded for interventional and clinical data, independently evaluated pre-TARE CT in a random order. For image segmentation, the reader-specific label map was created on overall liver tumor burden in portal venous phase CT images. For a better reliability, image de-noising was performed using wavelet transform. RF from labelled CT images were extracted twice, each by the same independent readers for inter-observer analysis, and included 162 first-order logic features and 216 gray level co-occurrence matrix (GLCM) features. These features are used to quantify tumor size (e.g., volume), shape (e.g., compactness, sphericity), and intensity (e.g., histogram statistics of mean, standard deviation, median), as well as texture matrices including the GLCM, where the differences represent the heterogeneity of the tumor.
Feature selection and model analysis
Feature selection and dimension reduction were necessary, as the number of RFs (n = 378) exceeded the number of patients (n = 51). The reproducibility of the extracted features between the two readers was assessed by calculating the concordance correlation coefficient (CCC) for each of the features as a measure of intra-class correlation. Features with a coefficient between 0.8 and 1 were classified as "excellent" and included in further analysis. Using z-score standardization, all feature values were normalized to a range between 0 and 1, which improves comparability. The normalized dataset was randomly subdivided into a balanced training and test dataset (70/30 ratio). Further feature reduction was performed only on the training dataset using a Boruta machine learning algorithm. Using the RFs determined to be reliable, a radiomics signature was constructed to predict the treatment response in the training dataset. Finally, a multivariate logistic regression analysis was performed on the test dataset to fit and test the model. The discriminatory efficacy of the features was quantified by calculating the area under the curve (AUC) using receiver operating characteristic (ROC) by applying a model-derived threshold. The radiomics workflow is illustrated in Figure 1.
Statistical Analysis
The characteristics age, sex, pre-therapeutic laboratory data, primary tumor side, metastatic tumor burden of the liver, and KRAS status were analyzed regarding their predictive value for therapy response at follow-up after 3 months with a binary logistic regression for metric and categorial predictors. A radiomics-based model was built based on a radiomics signature consisting of reliable RFs that allow classification of response using multivariate logistic regression. According to a cutoff determined in the model, patients were assigned to either high- or low-risk groups for disease progression after TARE. Kaplan-Meier analysis was performed to analyze survival between high- and low-risk groups. P values < 0.05 were considered to be statistically significant.