Best Practices in Liver Biopsy Histologic Assessment for Nonalcoholic Steatohepatitis Clinical Trials: Expert Opinion

Table 3

Artificial intelligence based and other new methodologies for quantitative assessment of features of NASH and liver fibrosis.

Brand name		Available literature
Histoindex		Astbury et al.; Liu et al.; Bedossa et al.; Wang et al.
Brief description	(i) Uses SHG/TPEF imaging-based tool to provide an automated, quantitative assessment of histological features pertinent to NASH (fibrosis and components of the NAS).; (ii) The generated data quantifies fibrosis, steatosis, ballooning, and inflammation. It provides measurements of disease progression and regression in NASH
Advantages	(i) Can stage samples as small as 0.5-1.0 cm; (ii) Stain-free imaging may reduce staining-related variation in interpretation; (iii) Reproducible qFfibrosis; prelim outcome data for HBV/NASH; (iv) Stain-free imaging enables co-localization for fibrosis, steatosis, ballooning, and inflammation, which are all obtained on the same slide, which was not possible using conventional methods using multiple staining from consecutive slides.; (v) AI-based algorithm is developed based with pathologists and is knowledge based, interpretation of the reading is logical for clinical and pathological assessment
Limitations	(i) Slow and expensive: (scanning takes ~2 hours per slide).; (ii) Cannot detect both faint and dense collagen simultaneously; (iii) Performance best for assessment of degree of steatosis and fibrosis (strong correlation with pathologists’ scores for qFibrosis () and qsteatosis () compared to severe inflammation and higher ballooning grades)
Biocellvia		DeRudder et al.; Albadrani et al.
Brief description	(i) Uses multiparametric image analysis and computerized analysis of high-resolution digitized whole histological sections.; (ii) Combines specific stains and immunohistochemistry with advanced algorithms to quantify key morphometric parameters: Ssteatosis, collagen fibers, inflammation, ballooning, ductular reaction; (iii) Whole section analyzed simultaneously
Advantages	(i) Short turnaround time (<1 h for 100 slides).; (ii) Quantified histology has been validated in preclinical models of liver and lung fibrosis; (iii) With superior performance compared with traditional scoring with respect to accuracy, reliability, reproducibility, and speed; (iv) Purported to eliminate variability in interpretation and provide better insight into a compound’s efficacy; (v) Whole section analysis; (vi) Provide zonal distribution of fibrosis (perisinusoidal, vascular, and septal) in human
Limitations	(i) No quantification of portal fibrosis; (ii) Need to check correlation for inflammation in human
PathAI		Taylor-Weiner et al.; Carrasco-Zevallos et al.
Brief description	(i) PathAI has developed a ML-based approach to liver histology assessment, characterizing disease severity and heterogeneity, and quantifying treatment response in NASH.; (ii) Deep convolutional neural networks are leveraged for identifying and characterizing NASH histologic features.
Advantages	(i) PathAI ML scores and quantitative features are generated on digitized H&E and tTrichrome slides evaluated by central readers, facilitating integration of ML approach into existing NASH clinical trial workflows.; (ii) ML-derived continuous quantitative histologic features capture subtle histologic changes not detectable through manual scoring alone; agreement between ML-derived and manual consensus scores for NASH is stronger than pathologist inter-rater agreement
Limitations	(i) The ML models are trained using inputs from NASH pathologists, who demonstrate high inter- and intrarater variability in scoring of NASH biopsies.; (ii) PathAI ML algorithms can be sensitive to variability in H&E or tTrichrome stain quality
Fibronest		Lara et al.
Brief description	(i) Translational quantitative image analysis for the quantification of fibrosis and associated histological features of NASH; (ii) FibroNest-pPredict uses AI to link digital pathology images and outcomes/biomarkers and establish image-based predictive models.
Advantages	(i) Quantifies same slide/image as used by pathologists to generate automated continuous scores and augmented pathology images to assess fibrosis severity and disease activity across multiple fibrotic conditions (liver, lung, kidney, skin, muscle, and heart).; (ii) Turnaround time approximately 2 weeks.; (iii) Availability of preclinical and translational supportive data
Limitations	(i) Young company; (ii) Recent involvement in pPhase 2-3 trials; (iii) Prospective data is pending

Abbreviations: AI = artificial intelligence; HBV = hepatitis B virus; H&E = hematoxylin and eosin; ML = machine learning; NAS = NAFLD activity score; NASH = nonalcoholic steatohepatitis; qFibrosis = quantitative assessment of liver fibrosis; SHG = second harmonic generation; TPEF = two-photon excitation fluorescence.