Room: Exhibit Hall | Forum 8
Purpose: The aim of this study is to evaluate diagnostic performance of a deep learning scheme having as input ultrasound (US) B-Mode images containing right kidney cortex and liver parenchyma. Those areas are indicated by an expert radiologist, on patients with Non-Alcoholic Fatty Liver Disease (NAFLD) using liver biopsy (LB) as â€˜Gold Standardâ€™.
Methods: 112 consecutive NAFLD diagnosed patients (37 S0, 43 S1, 25 S2 and 7 S3) underwent a regular abdomen US B-Mode and LB examination employing the Kleiner score (S0-S3). A B-Mode image containing right kidney cortex (RK) and liver parenchyma (LP) was acquired and saved by an expert Radiologist performing the US examination. Then, the radiologist indicated a point on the surface separating RK and LP for each patient on the saved image and a 225x225 window around that point was cropped automatically. The cropped images were fed to the AlexNet pre-trained deep learning scheme using the LB steatosis grading (S0-S3) as labels. Data were separated in two classes according to their steatosis grade: S0-S1 (Mild Steatosis Class) as class 1, and S2-S3 (Significant Steatosis Class) as class 2. The images were separated in training (70%) and testing (30%) and corresponding training and testing accuracies were calculated.
Results: The deep learning scheme reached 100% accuracy for the training data while the accuracy of the trained model on the test data reached 82.35%.
Conclusion: The results indicate that this scheme can be used as a supplementary tool for NAFLD diagnosis and especially for significant steatosis detection.