Deep learning helps discriminate between autoimmune hepatitis and primary biliary cholangitis

Abstract

Background & aims: Biliary abnormalities in autoimmune hepatitis (AIH) and interface hepatitis in primary biliary cholangitis (PBC) occur frequently, and misinterpretation may lead to therapeutic mistakes with a negative impact on patients. This study investigates the use of a deep learning (DL)-based pipeline for the diagnosis of AIH and PBC to aid differential diagnosis.

Methods: We conducted a multicenter study across six European referral centers, and built a library of digitized liver biopsy slides dating from 1997 to 2023. A training set of 354 cases (266 AIH and 102 PBC) and an external validation set of 92 cases (62 AIH and 30 PBC) were available for analysis. A novel DL model, the autoimmune liver neural estimator (ALNE), was trained on whole-slide images (WSIs) with H&E staining, without human annotations. The ALNE model was evaluated against clinico-pathological diagnoses and tested for interobserver variability among general pathologists.

Results: The ALNE model demonstrated high accuracy in differentiating AIH from PBC, achieving an area under the receiver operating characteristic curve of 0.81 in external validation. Attention heatmaps showed that ALNE tends to focus more on areas with increased inflammation, associating such patterns predominantly with AIH. A multivariate explainable ML model revealed that PBC cases misclassified as AIH more often had ALP values between 1 × upper limit of normal (ULN) and 2 × ULN, coupled with AST values above 1 × ULN. Inconsistency among general pathologists was noticed when evaluating a random sample of the same cases (Fleiss's kappa value 0.09).

Conclusions: The ALNE model is the first system generating a quantitative and accurate differential diagnosis between cases with AIH or PBC.

Impact and implications: This study demonstrates the significant potential of the autoimmune liver neural estimator model, a transformer-based deep learning system, in accurately distinguishing between autoimmune hepatitis and primary biliary cholangitis using digitized liver biopsy slides without human annotation. The scientific justification for this work lies in addressing the challenge of differentiating these conditions, which often present with overlapping features and can lead to therapeutic mistakes. In addition, there is need for quantitative assessment of information embedded in liver biopsies, which are currently evaluated on qualitative or semi-quantitative methods. The results of this study are crucial for pathologists, researchers, and clinicians, providing a reliable diagnostic tool that reduces interobserver variability and improves diagnostic accuracy of these conditions. Potential methodological limitations, such as the diversity in scanning techniques and slide colorations, were considered, ensuring the robustness and generalizability of the findings.

Keywords: Artificial intelligence; Autoimmunity; Computational pathology; Digital pathology; Liver; Rare liver diseases.

Graphical abstract

Fig. 1

Schematic of the workflow and results for the ALNE model. (A) Graphical representation of the workflow for classification of PBC and AIH on histopathology WSI data. The process involves transformer-based feature extraction and MIL techniques for effective classification and visualization. (B) Three-fold cross-validated AUROC scores for classifying PBC and AIH on the internal validation cohort utilizing the ALNE model. (C) AUROC scores, based on 10 repetitions, for the classification of PBC and AIH in the external validation cohort using the ALNE model. AIH, autoimmune hepatitis; ALNE, autoimmune liver neural estimator; AUROC, area under the receiver operating characteristic curve; MIL, multiple instance learning; PBC, primary biliary cholangitis; WSI, whole-slide image.

Fig. 2

Visualization of the attention heatmaps for the validation cohort. (A) Correctly predicted instances. Above: terminal hepatic vein with adjacent mild sinusoidal dilatation and no inflammation, indicating PBC. Below: portal tract exhibiting moderate chronic inflammation and interface activity, correctly classified as AIH. (B) Incorrectly predicted instances. Above: parenchymal area without necroinflammatory foci or confluent necrosis, misclassified as PBC. Below: moderate chronic inflammation in an enlarged portal tract with lymphocytic cholangitis and degenerative changes of the bile duct, misclassified as AIH. AIH, autoimmune hepatitis; PBC, primary biliary cholangitis.

Fig. 3

Evaluating pathologist interobserver variability and AI conformity. The paired agreement among couples of pathologists (at the top) and between each pathologist and the AI model (at the bottom) in classifying AIH and PBC is represented by a variation of the Cohen’s kappa index. The Cohen’s kappa index is a metric which runs between -1 and 1 and takes into account agreement by chance. The subplot at the top shows the agreement between each pair of pathologists, whereas the agreement between the AI model and each pathologist is shown in the subplot at the bottom. The analysis was performed on a random subset of 19 cases from the validation cohort. For evaluation purposes, the pathologists assessed each case using the H&E slides only. AI, artificial intelligence; AIH, autoimmune hepatitis; PBC, primary biliary cholangitis.