High performance deep-learning model for the diagnosis of auto-immune hepatitis based on histological whole slide images.

Autoimmune Hepatitis (AIH) is a liver disease with a wide clinical spectrum, driven by an abnormal immune response against the liver parenchyma. Challenges persist especially in terms of accurate diagnosis of acute onsets and differential diagnosis, outlined by the European Association for the Study of Liver guidelines. Our training cohort comprised 170 untreated AIH and 232 control cases with a variety of differential diagnosis. Ground Truth was the integrative diagnosis of the clinical, histological, biological and treatment response data. We trained a multiple-instance deep-learning model (DLM) from Whole Slide Images alone for the initial diagnosis of AIH. The model was then tested on an external dataset of 61 AIH and 124 controls. Prospective real-life testing of the model was conducted between January and June 2025. Our DLM "Artificial Intelligence On Liver Immunity (AIOLI)" achieved an AUC of 0,92 ± 0,02 on the training dataset and an AUC of 0,74 on the external dataset; in detail, it achieved an AUC of 0,89 for the differential diagnosis of AIH vs. acute alcoholic hepatitis, 0,98 vs. MASH, 0,42 vs. Hepatitis B Virus and 0,76 vs. drug-induced liver injuries. Retrieval of the five most predictive tiles allowed to identify patterns used by the model for prediction and provided interpretability. In the prospective setting, AIOLI achieved a Sensibility of 0.86, a Specificity of 0.76, a F1-score of 0.69 and an AUC of 0.73. AIOLI is an interpretable DLM able to segregate AIH from a variety of control cases with performances comparable to those of an expert liver pathologist, setting a benchmark for future research. We plan to enhance our performances by enriching our dataset, and to validate this approach with a multicentric deployment.