Abstracts
 
RETURN TO ABSTRACT LISTING


2832. Accuracy of Machine Learning-Based Radiomics for Differentiating Diffuse Liver Diseases on Non-Contrast CT
Authors * Denotes Presenting Author
  1. Fatemeh Homayounieh; Harvard Medical School; Massachusetts General Hospital
  2. Leila Mostafavi *; Harvard Medical School; Massachusetts General Hospital
  3. Sanjay Saini; Massachusetts General Hospital
  4. Ruhani Doda Khera; Harvard Medical School; Massachusetts General Hospital
  5. Bernhard Schmidt; Siemens Healthcare GmbH
  6. Thomas Flohr; Siemens Healthcare GmbH
  7. Mannudeep Kalra; Harvard Medical School; Massachusetts General Hospital
Objective:
From the subjective semantic interpretation of images and findings, radiomics move cross-sectional imaging into the domain of quantitative imaging with several mathematical features which help assess the lesions, their stoma, as well as in their temporal monitoring. Although benign conditions such as hepatic fibrosis and non-alcoholic steatohepatitis have been assessed with radiomics, most radiomics studies focus on the oncologic applications of cross-sectional imaging from the perspective of lesion characterization, mutation type, and prediction of outcome or treatment response [1-6]. In this context, machine learning algorithms offer an opportunity to simplify, automate, and integrate radiomics into the routine clinical practice. Our study assesses the accuracy of machine learning (ML)-based radiomics for differentiating healthy liver from diffuse liver diseases (cirrhosis, steatosis, amiodarone toxicity, and iron overload) on non-contrast abdomen CT.

Materials and Methods:
Our IRB-approved study included 300 adult patients (mean age 63±16 years; 171 men, 129 women) who underwent non-contrast abdomen CT and had either a healthy liver (n=100 patients) or an evidence of diffuse liver disease (n=200). The diffuse liver diseases included steatosis (n=50), cirrhosis (n=50), hyperdense liver due to amiodarone toxicity (n=50) or iron overload (n=50). All patients had an abdominal MRI or a history compatible with their findings. De-identified image sets were post-processed on an offline Radiomics prototype. Semiautomatic segmentation of liver was performed in one section at the level of the porta hepatis (all 300 patients), and then over the entire liver volume (50 patients). The prototype estimated radiomics features within the segmented portion and performed statistical comparison of healthy and abnormal liver with multiple logistic regression tests and random forest classifier.

Results:
With random forest classifier, the area under the curve (AUC) for radiomics ranged between 0.72 (iron overload vs. healthy liver) to 0.98 (hepatic steatosis vs. healthy liver) for differentiating diffuse liver disease from the healthy liver. Combined root-mean-square and gray level co-occurrence matrix (GLCM) demonstrated the highest AUC (AUC:0.99, p<0.01) for differentiating healthy liver from steatosis. These features plus small area low gray level emphasis were most accurate differentiators of healthy liver and cirrhosis (AUC:0.88, p<0.0004). Radiomics were more accurate for differentiating healthy liver from amiodarone (AUC:0.93) than from iron overload (AUC:0.79).

Conclusion:
Machine learning-based radiomics of unenhanced abdomen CT enable differentiation of healthy liver from hepatic steatosis, cirrhosis, amiodarone toxicity, and iron overload.