Precise, quantitative assessment of disease features on histological images from preclinical models is essential for therapeutic development in diseases such as pulmonary fibrosis. However, current histological fibrosis scoring methods, such as Ashcroft Scoring have several limitations, including high time and labor requirements by expert pathologist readers, subjective semi-quantitative assessment and interobserver variability. The goal of this study was to assess the feasibility of a supervised AI/ML-based framework for automated, rapid, objective quantification of fibrosis content and spatial distribution in histology slides obtained from a variety of preclinical fibrosis models. Lung histology slides stained with Masson’s trichrome were obtained from 194 individual mice from two independent cohorts of preclinical mouse models of pulmonary fibrosis. A supervised AI/ML algorithm was trained, validated and independently tested to automatically detect, segment and quantify fibrosis compared against independent Ashcroft scoring by an expert pathologist reader. Spatial distribution of AI/ML-segmented fibrosis patterns were compared across histology images. AI/ML-based fibrosis quantification demonstrated strong correlation with Ashcroft score, both in the validation cohort (Spearman ρ = 0.85, CI: 0.72–0.92), and in the independent, de novo test cohort (Spearman ρ = 0.89, CI: 0.84–0.93) with rapid assessment time (~ 1.5 times faster). Additionally, Ripley’s K analysis revealed differences in spatial distribution of AI-segmented fibrosis patterns among samples with similar Ashcroft scores and overall fibrosis content. The AI/ML framework developed and independently validated in this study provides a robust, computationally-efficient method for precise, user-friendly, objective measurement of fibrosis content and spatial distribution, which would have major utility in preclinical therapeutic trials and investigations of disease pathogenesis.
Roop et al. (Wed,) studied this question.