ARRS 2022 Abstracts


2036. Radiomic Tumor Features from MRI for Non-Invasive HCC Grading
Authors * Denotes Presenting Author
  1. Sanaz Ameli; University of Arkansas for Medical Sciences
  2. Bharath Ambale Venkatesh; Johns Hopkins Hospital
  3. Mohammadreza Shaghaghi *; Johns Hopkins Hospital
  4. Bita Hazhirkarzar; Johns Hopkins Hospital
  5. Roya Rezvani Habibabadi; University of Florida
  6. Mounes Aliyari Ghasabeh; St. Louis University
  7. Ihab Kamel; Johns Hopkins Hospital
This study aims to evaluate the ability of radiomics in hepatocellular carcinoma (HCC) differentiation based on apparent diffusion coefficient (ADC) and venous enhancement from MRI.

Materials and Methods:
This retrospective study was HIPPA compliant and IRB approved. A total of 98 HCC lesions with baseline MRI and pathologic reports were identified between January 2001 and June 2017. Volumetric measurements of venous enhancement (VE) and ADC were performed on baseline MRI. The tumors were histologically classified into well and poorly differentiated. A MATLAB-based program was used to perform texture analysis, which extracted 95 texture features from the tumor volumes, 46 each for the ADC and VE maps, and three shape-related. Once the features were extracted using texture analysis, a statistical machine learning technique, random forest classification was used to identify the performance of the combined set of texture features in classification of tumors into well and poorly differentiated (based on the pathology report). The Gini index was used for split criterion, the random forest model was optimized to have a minimum of two participants per leaf node, and 10 randomly chosen variables were used at each split. The number of decision trees grown were 1000. Predictor importance was estimated based on the out-of-bag samples by permutation. If a predictor is influential in prediction, then permuting its values should affect the model error. If a predictor is not influential, then permuting its values should have little to no effect on the model error.

A total of 98 HCC lesions were evaluated. Out of 95 radiomic features, the top four were ADC features. The overall model with all features of HCC tumors had an AUC of 81%. ADC features were superior to VE features for tumor classification. The overall error rate in tumor classification was 23.4%. The error rates in classifying well and poorly differentiated HCCs were 26% and 21%, respectively.

Radiomics might be a useful technique for non-invasive tumor differentiation without the need for pathology. Out of all radiomic features, ADC features showed superior performance in classifying HCC tumors as compared to VE features.