Smith, Lindsey

AimThe aim of this study was to compare the histopathologic features of MMF‐induced colitis and colonic GVHD and develop a digital tool using deep learning convolutional neural networks (CNNs) to semi‐automate the quantification of eosinophils.MethodsMMF and GVHD colitis cases were retrieved. A cohort of transplant patients who were at risk for developing GVHD and were on MMF at the time of biopsy was also identified. Cases were reviewed for histologic features and a deep learning CNN was trained to quantify lamina propria eosinophils. Clinical history was collected.ResultsNinety‐five patients were included in the study group: GVHD (n = 37), MMF (n = 25) and GVHD vs. MMF (n = 33). Within the GVHD vs. MMF group, one patient was thought to have definitive MMF colitis and one possible MMF colitis. The GVHD group had a significantly higher apoptotic count compared to the MMF group (median: 7 vs. 2, P = 0.03). Neuroendocrine aggregates were present in the GVHD (N = 12) and GVHD vs. MMF (N = 4) groups while absent in the MMF group. The MMF group had a higher semi‐quantitative eosinophil score and a higher automated eosinophil count as detected by the deep learning CNN (median eos/HPF: MMF 19.2, GVHD vs. MMF 3.6, GVHD 3.4, P < 0.001).ConclusionsLower eosinophil counts, higher apoptotic counts and neuroendocrine cell aggregates favour the diagnosis of GVHD. AI models for eosinophil quantification may serve as an ancillary tool for assessing this differential diagnosis. Given the rarity of MMF‐induced colitis in SCT patients, pathologists should have a high threshold for making this diagnosis in this population.

Aims Reticulin stain is used routinely in the histological evaluation of hepatocellular carcinoma (HCC). The goal of this study was to assess whether the histological reticulin proportionate area (RPA) in HCCs predicts tumour‐related outcomes. Methods and results We developed and validated a supervised artificial intelligence (AI) model that utilises a cloud‐based, deep‐learning AI platform (Aiforia Technologies, Helsinki, Finland) to specifically recognise and quantify the reticulin framework in normal livers and HCCs using routine reticulin staining. We applied this reticulin AI model to a cohort of consecutive HCC cases from patients undergoing curative resection between 2005 and 2015. A total of 101 HCC resections were included (median age = 68 years, 64 males, median follow‐up time = 49.9 months). AI model RPA reduction of > 50% (compared to normal liver tissue) was predictive of metastasis [hazard ratio (HR) = 3.76, P = 0.004, disease‐free survival (DFS, HR = 2.48, P < 0.001) and overall survival (OS), HR = 2.80, P = 0.001]. In a Cox regression model, which included clinical and pathological variables, RPA decrease was an independent predictor of DFS and OS and the only independent predictor of metastasis. Similar results were found in the moderately differentiated HCC subgroup (WHO grade 2), in which reticulin quantitative analysis was an independent predictor of metastasis, DFS and OS. Conclusion Our data indicate that decreased RPA is a strong predictor of various HCC‐related outcomes, including within the moderately differentiated subgroup. Reticulin, therefore, may represent a novel and important prognostic HCC marker, to be further explored and validated.