Abstracts

RETURN TO ABSTRACT LISTING


1693. Pretreatment CT Radiomics Predicts Response to Chemotherapy in Patients with Colorectal Liver Metastasis Undergoing Curative-Intent Surgery
Authors * Denotes Presenting Author
  1. Michael Markovitz *; University of South Florida
  2. Justin Wilkes; Moffitt Cancer Center
  3. Masoumeh Ghayouri; Moffitt Cancer Center
  4. Jason Denbo; Moffitt Cancer Center
  5. Jasmina Ehab; University of South Florida
  6. Daniel Anaya; Moffitt Cancer Center
  7. Daniel Jeong; Moffitt Cancer Center
Objective:
To evaluate the capacity of pre-chemotherapy computed tomography (CT) radiomics in predicting post treatment response, using histologic tumor regression grade (TRG) in surgical colorectal metastasis hepatectomy specimens.

Materials and Methods:
75 subjects (32 females, age 56.7±10.6 years) with colorectal adenocarcinoma liver metastases receiving preoperative chemotherapy and subsequent partial liver resection (7/2015 to 3/2020) were included in this retrospective cohort. On pre-chemotherapy axial CT venous phase post contrast images, the dominant hepatic metastasis was segmented on each slice to generate a volume of interest (VOI). 572 radiomics features were extracted using Healthmyne (Healthmyne Inc, Madison, WI) from the VOIs. Corresponding histologic TRG was measured on surgical pathology specimens and patients were classified as responders (TRG 1-2, TRGr) or non-responders (TRG 3-5, TRGnr). Response evaluation criteria in solid tumors (RECIST) v1.1 was also measured using pre and post-chemotherapy/pre-surgical CT exams [1]. Correlation analysis and logistic regression models were performed to predict non-responders (TRGnr) using pretreatment CT radiomics. Categorical variables were compared using Chi Square and continuous variables were compared using Mann Whitney U. The accuracy of the models was examined with receiver operating characteristics (ROC) analysis.

Results:
Twenty-three radiomics features differentiated TRGr from TRGnr (p<0.05) in univariate analysis, including size (n=5), shape (n=1), density (n=1), and texture (n=16) features. Gray Level Co-occurrence Matrix (GLCM) correlation showed the highest univariate diagnostic accuracy with an Odds Ratio of 1.92 (1.12-3.27; 95% CI) and AUC of 0.69 (0.56-0.82; 95% CI), p=0.02. A multivariable model of 8 non-correlating features yielded an AUC of 0.72 (0.60-0.84; 95% CI), p=0.002. Of note, RECIST v1.1 response class using both pre- and post-chemotherapy CT exams was not significant in determining TRGr vs. TRGnr, p=0.26.

Conclusion:
Colorectal adenocarcinoma liver metastases are commonly treated with neoadjuvant chemotherapy with the goal of improving surgical resectability and overall outcome. However, response to chemotherapy is variable and to some extent unpredictable. This noninvasive tool to predict who will benefit from neoadjuvant therapies could be beneficial in risk-stratification and in guiding therapy. Our combined model achieved a diagnostic accuracy of 72% in predicting histologic TRG after chemotherapy on pretreatment CT. Future larger and prospective studies with clinical data integration may further elucidate the role CT radiomics could play in clinical treatment pathways.