Digital Hepatic Iron Content: An Artificial Intelligence Model for Spatially Resolved Histologic Iron Quantitative Analysis in Liver Samples

• Published in: Laboratory investigation Vol. 103; no. 9; p. 100200
• Main Authors: Sivasubramaniam, Priyadharshini, Stokes, Nadarra, Patil, Ameya, Smith, Lindsey, Hartley, Christopher P., Graham, Rondell P., Moreira, Roger K.
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01-09-2023
• Subjects: digital pathology; digital quantification; hepatic iron; hepatic iron index; iron index; Prussian blue; hepatic iron; hepatic iron index; Prussian blue; iron index; digital quantification; digital pathology
• ISSN: 0023-6837; 1530-0307
• DOI: 10.1016/j.labinv.2023.100200

Currently, the precise evaluation of tissue hepatic iron content (HIC) requires laboratory testing using tissue-destructive methods based on colorimetry or spectrophotometry. To maximize the use of routine histologic stains in this context, we developed an artificial intelligence (AI) model for the recognition and spatially resolved measurement of iron in liver samples. Our AI model was developed using a cloud-based, supervised deep learning platform (Aiforia Technologies). Using digitized Pearl Prussian blue iron stain whole slide images representing the full spectrum of changes seen in hepatic iron overload, our training set consisted of 59 cases, and our validation set consisted of 19 cases. The study group consisted of 98 liver samples from 5 different laboratories, for which tissue quantitative analysis using inductively coupled plasma mass spectrometry was available, collected between 2012 and 2022. The correlation between the AI model % iron area and HIC was Rs = 0.93 for needle core biopsy samples (n = 73) and Rs = 0.86 for all samples (n = 98). The digital hepatic iron index (HII) was highly correlated with HII > 1 (area under the curve [AUC] = 0.93) and HII > 1.9 (AUC = 0.94). The percentage area of iron within hepatocytes (vs Kupffer cells and portal tract iron) identified patients with any hereditary hemochromatosis–related mutations (either homozygous or heterozygous) (AUC = 0.65, P = .01) with at least similar accuracy than HIC, HII, and any histologic iron score. The correlation between the Deugnier and Turlin score and the AI model % iron area for all patients was Rs = 0.87 for total score, Rs = 0.82 for hepatocyte iron score, and Rs = 0.84 for Kupffer cell iron score. Iron quantitative analysis using our AI model was highly correlated with both detailed histologic scoring systems and tissue quantitative analysis using inductively coupled plasma mass spectrometry and offers advantages (related to the spatial resolution of iron analysis and the nontissue-destructive nature of the test) over standard quantitative methods.