Self-Supervised Learning Model Accurately Diagnoses Hemophagocytic Lymphohistiocytosis from Histology Images

Background Hemophagocytic lymphohistiocytosis (HLH) is a rare multi-system inflammatory disorder characterized by hyperactive macrophages that phagocytize blood cells. The diagnosis of HLH is based on meeting at least 5 or 9 diagnostic criteria, of which morphologic evidence of hemophagocytosis is one. Morphologic assessment of HLH is based on evaluation of bone marrow aspirates or biopsies; however, evaluation of BMBx can be challenging due to the frequent subtle histologic changes seen in HLH, as well as the fact the hemophagocytosis may be seen in association with other hematologic disorders. To address this diagnostic challenge, we investigated whether machine learning methods utilizing a self-supervised learning algorithm can distinguish HLH from its clinical mimics based on bone marrow biopsy images. Methods We employed a self-supervised approach using the Histomorphological Phenotype Learning (HPL) with a BarlowTwins model that autonomously identifies unique features in tiled biopsy images. Our dataset comprised biopsy images collected from NYU Langone Hospital between 2014 and 2023, including 32 patients with HLH, as well as a large number of non-HLH patients with diagnoses of myeloproliferative neoplasms (n=53), negative lymphoma staging BMBx (n=100), myelodysplastic syndrome (n=89), and non-HLH or MDS-related cytopenias (n=55). Whole slide images (WSI) of BMBx's were divided into 224px tiles at 20x magnification and input into the model. The output was grouped into clusters and visualized using Uniform Manifold Approximation and Projection (UMAP) that denoted the distribution of clusters with regard to the phenotypes they were grouped by. The classification performance of the model was assessed using a logistic regression algorithm with significant clusters. The evaluation utilized the area under the receiver operating characteristic curve (AUROC) metric to determine the effectiveness of the algorithm in differentiating HLH from the non-HLH patient groups. Results We identified 45 tile clusters based on aggregate analysis of all samples. Among these, HLH cases were enriched for one cluster and depleted for seven. The single tile cluster enriched in HLH included images that captured macrophages visibly phagocytosing cells. The eight clusters enriched or depleted in HLH were used for logistic regression analysis. A logistic regression model demonstrated high performance in distinguishing HLH from other diseases. For the HLH versus others scenario, the overall AUROC was 0.95 with a standard deviation of ±0.05. Specific comparisons yielded the following results: HLH versus negative lymphoma staging controls showed a mean AUROC of 0.90 ±0.07. HLH versus non-HLH/MDS cytopenic controls achieved a mean AUROC of 0.98 ±0.04. HLH versus MDS resulted in a mean AUROC of 0.88 ±0.08. Lastly, HLH versus MPN showed a mean AUROC of 0.95 ±0.05. Conclusion Our study demonstrates the potential of using self-supervised learning algorithms to identify distinct features in biopsy images to diagnose HLH. This method could significantly improve the accuracy and efficiency of HLH diagnosis with visually featured pattern tiles, reducing reliance on expensive and tedious conventional methods.