Abstracts
 
RETURN TO ABSTRACT LISTING


2984. Can Machine Learning Radiomics Predict and Measure Treatment Response in Metastatic Liver Disease?
Authors * Denotes Presenting Author
  1. Leila Mostafavi *; Harvard Medical School; Massachusetts General Hospital
  2. Fatemeh Homayounieh; Harvard Medical School; Massachusetts General Hospital
  3. Subba Digumarthy; Harvard Medical School; Massachusetts General Hospital
  4. Gina Basinsky; Harvard Medical School; Massachusetts General Hospital
  5. Gordon Harris; Harvard Medical School; Massachusetts General Hospital
  6. Mannudeep Kalra; Harvard Medical School; Massachusetts General Hospital
Objective:
Recent studies have explored the role of radiomics for predicting treatment response in several malignancies, including liver metastases [1,2]. Radiomics is used to predict treatment response in subjects hepatocellular and esophagogastric cancers [1,2]. We hypothesized that radiomics can predict and measure treatment response in patients with metastatic liver disease from breast cancer. The purpose of our study was to assess if radiomics can predict and evaluate treatment response in patients with metastatic breast cancer.

Materials and Methods:
Our IRB approved study included 203 women (mean age 54±11 years) with metastatic liver disease from breast cancer. All patients underwent contrast abdomen-pelvis computed tomography (CT) in portal venous phase at two timepoints - baseline (BL: pre-treatment) and follow-up (FU: between 3-12 months following treatment). Patients were subcategorized into three subgroups based on RECIST 1.1. criteria (Response Evaluation Criteria in Solid Tumors version 1.1): 66 with stable disease (SD), 69 with partial response (PR) and 68 with progressive disease (PD) on follow-up CT. CT images from BL and FU were deidentified and exported to radiomics prototype (eXamine, Siemens Healthineers). The prototype enabled semiautomatic segmentation of the target liver lesions for extraction of first and high order radiomics. Statistical analyses with logistic regression and random forest classifiers was performed with the prototype to assess how well BL radiomics predicts treatment response, and whether radiomics can differentiate SD from PD and PR on the two timepoints.

Results:
There was no significant difference between the radiomics on baseline and follow-up CT images of patients with SD (area under the curve (AUC) 0.3). Random forest classifier differentiated patients with PR on the baseline and follow-up images with an AUC of 0.845. The top most relevant features on forest plots were high and low pass wavelet filters of the large dependence emphasis (derived gray level dependence matrix features (GLDM) features), dependence variance (an original GLDM feature), and flatness (a shape-based radiomic). Random forest classifier differentiated PD on baseline and follow-up images with an AUC of 0.731. The top most relevant features on random forest classifiers were surface to volume ratio (original shape feature), followed by low pass wavelet filter of informational measure of correlation 2 - IMC2, and high pass wavelet filter of zone entropy.

Conclusion:
Machine learning-based semiautomatic segmentation and radiomics prototype enable differentiation of stable disease from a partial response and a progressive metastatic liver disease on baseline and follow-up CT. Despite substantial variations in the scanners, acquisition, and reconstruction parameters in our study, radiomics had an AUC of 0.84-0.89 for differentiating stable hepatic metastases from the decreasing and increasing metastatic disease. On the other hand, the prediction of disease progression or partial response from the baseline CT was not successful with the radiomics (AUC 0.53-0.58).