Deep-Learning-Aided Quantification of Steatohepatitis-Associated Pathological Findings in Liver Specimens.

Background: Although semi-quantitative scoring is a common method for quantifying histopathological findings, semi-quantitative evaluation may lead to errors in objectivity and reproducibility depending on the histological findings. We developed a deep learning model to quantify steatohepatitis-associated pathological findings from liver specimens.

Methods: Eighteen liver specimens of steatohepatitis, and eight liver specimens of chronic hepatitis were used to train AI (convolutional neural network), and construct an AI to quantify steatohepatitis-associated pathological findings. The AI model measured percentages of the steatosis area (%Steatosis), ballooning area (%Ballooning), and fibrosis area (%Fibrosis) in the liver biopsy specimens. Subsequently, 233 patients who underwent ultrasonography before liver biopsy were used for validation. Histological findings of these liver specimens were evaluated by three evaluators with expertise in liver pathology.

Results: %Steatosis, %Ballooning, and %Fibrosis were measured for all liver specimens. In all patients, %Steatosis was strongly correlated with the histological steatosis grade (R = 0.78, p < 0.001) and ultrasound controlled attenuation parameter (R = 0.51, p < 0.001). In the liver tissues of 47 nonalcohol-related steatotic liver disease (NASLD) patients, %Ballooning was correlated with the histological ballooning grade (R = 0.4, p < 0.01). In the NASLD patients, %Fibrosis was strongly correlated with the histological fibrosis stage (R = 0.72, p < 0.001), and ultrasound shear wave elastography (R = 0.7, p < 0.001).

Conclusion: Deep-learning-aided pathology can quantify steatohepatitis-associated pathological findings from liver tissue specimens. This AI model is a tool used to quantify the steatohepatitis-associated pathological findings and is not an algorithm to diagnose steatohepatitis.