A Deep Learning Trial on Transient Elastography for Assessment of Liver Fibrosis

Transient elastography (TE) is a non-invasive, rapid and reproducible technology, which performs liver stiffness measurement (LSM) for staging the liver fibrosis in the clinic. In the procedure of LSM, an M-mode strain image (MSI) is estimated from the M-mode radiofrequency (RF) data. On the hypothesis that the MSI contains richer information than LSM, we propose a deep learning (DL) method based on the MSI to improve the performance of fibrosis staging. A multicenter study was conducted where both TE and liver biopsy were performed on 421 patients and finally 245 patients with 2,713 MSIs were qualified. An optimal deep learning model with 8 layers (5 convolutional layers and 3 fully-connected layers) was built and evaluated on the dataset. LSMs were also used to assess liver fibrosis for comparison, while liver biopsy acted as the gold standard. The receiver operating characteristic (ROC) curve was plotted to evaluate the performance for assessing significant fibrosis (>= F2), advanced fibrosis (>= F3) and cirrhosis (F4). The AUC of the DL method was 0.850 for >= F2. The AUCs of the DL method were 0.948 for >= F3 and 0.934 for F4, respectively, slightly better than those of the LSM method (0.935 for >= F3 and 0.908 for F4, respectively). To conclude, the DL method performs better than the LSM method for advanced fibrosis and cirrhosis and could be used in transient elastography for assessment of liver fibrosis.