The current gold standard for the diagnosis and assessment of non-alcoholic fatty liver disease (NAFLD) progression is liver biopsy, which is then followed by microscopic analysis by a pathologist.To assess disease progression, the Kleiner and Brunt scoring system is typically used. In short, this approach assesses three histological features of liver injury (ballooning, inflammation, and steatosis). Analyses of these features are then combined to generate a NAFLD activity score. This approach possesses three main limitations. Firstly, its resolution is limited, as it assesses features characterised by continuous biological states. Secondly, it requires the presence of a trained pathologist, an in-demand profession. Lastly, its results are associated with reduced precision and reproducibility: there is significant intra- and inter-operator variability between scores obtained by different pathologists.
The aim of this study was to showcase an alternative approach to assess NAFLD progression: it presents an open source, Artificial Intelligence (AI) solution to obtaining Kleiner scores which addresses the limitations of the traditional pathologist-dependent scoring technique. Results from a classical computer analysis (collagen area based on the colour) and deep-learning-generated fibrosis scores for 467 clinically indicated human liver biopsies were compared to those determined by pathologists. Both ‘classical image analysis’ and AI-generated scores are in the same numeric range as the features of the Kleiner and Brunt score (e.g., 0–4 for fibrosis), but with continuous scaled outputs for increased precision.
Classical image analysis was unable to distinguish controls (score 0 fibrosis) from score 1 fibrosis, as the latter’s features do not significantly affect the collagen area. However, it was able to discriminate higher fibrosis levels: stage 4 fibrosis (cirrhosis) markedly impacted collagen areas. In contrast, AI-fibrosis scores could differentiate initial fibrosis scores. This ability became pronounced as fibrosis stage increased. This finding highlights AI’s ability to recognize and quantify subtle fibrotic changes with high resolution and reproducibility. Despite this higher resolution, however, significance levels of AI-fibrosis scores were lower than those obtained via classical image analysis.
The high resolution and reproducibility of AI-based scoring for the assessment of NAFLD progression suggest that it may be a promising alternative to current pathologist-dependent approaches. Future studies into AI for the diagnosis and assessment of NAFLD should seek to extend its potential by investigating other features of interest (eg microsteatosis) or capturing isolated morphological alterations associated with fibrosis.