Interpretability and generalizability of a one-dimensional convolutional neural network method for hepatic steatosis characterization

Abstract

Nonalcoholic fatty liver disease (NAFLD) affects 25% of the population globally. We developed a one-dimensional convolutional neural network (1-D CNN) method for the ultrasound tissue characterization using radio-frequency (RF) data and demonstrated its potential for the liver fat classification and quantification. We investigate herein the interpretability and generalizability of the method to understand why it works, and whether the performance is affected by settings, transducers, and platforms. We studied under various conditions the performances of 1-D CNN for predicting steatosis using magnetic resonance imaging-estimated proton density fat fraction (MRI-PDFF) as reference (steatosis: PDFF > 5%). Three datasets were used, each containing ultrasound RF data acquired from adults and same-day MRI-PDFF estimates: (1) 200 normal and NAFLD participants scanned using the Siemens S2000 ultrasound system with the 4C1 transducer (1–4 MHz); (2) 87 participants with known/suspected NAFLD scanned using Siemens S3000 with the 4C1 and/or 6C1HD transducers (1.5–6 MHz); and (3) 46 participants with known/suspected NAFLD scanned using GE Logiq e9 with the C1-6 transducer (1–6 MHz). The 1-D CNN method is generalizable among various instrumentation settings (e.g., focal depth, time gain compensation), betweentransducers of similar frequencies, between platforms, but not between the fundamental and tissue harmonic image modes. [Work supported by R01DK106419.] Nonalcoholic fatty liver disease (NAFLD) affects 25% of the population globally. We developed a one-dimensional convolutional neural network (1-D CNN) method for the ultrasound tissue characterization using radio-frequency (RF) data and demonstrated its potential for the liver fat classification and quantification. We investigate herein the interpretability and generalizability of the method to understand why it works, and whether the performance is affected by settings, transducers, and platforms. We studied under various conditions the performances of 1-D CNN for predicting steatosis using magnetic resonance imaging-estimated proton density fat fraction (MRI-PDFF) as reference (steatosis: PDFF > 5%). Three datasets were used, each containing ultrasound RF data acquired from adults and same-day MRI-PDFF estimates: (1) 200 normal and NAFLD participants scanned using the Siemens S2000 ultrasound system with the 4C1 transducer (1–4 MHz); (2) 87 participants with known/suspected NAFLD scanned using Siemens ...