A Self-Supervised Machine Learning Model Can Differentiate MDS from Clinical Mimics Based on Bone Marrow Biopsy Features

Introduction:

The pathologic evaluation of bone marrow biopsies (BMBx) for diagnosing myelodysplastic syndrome (MDS) is challenging due to the limited cytologic detail provided by H&E sections. Diagnosing MDS requires a sophisticated, integrative approach because of its complex clinical and histopathological presentations. In this study, we investigate whether integrating features from both H&E and reticulin-stained BMBx slides using self-supervised learning (SSL) image analysis could address these limitations by developing an algorithm capable of reliably distinguishing MDS from its clinically relevant mimics.

Methods:

In this study, four cohorts were included: 1) 89 patients with a verified diagnosis of MDS (“MDS” group); 2) 55 cytopenic patients who were suspected of MDS but did not meet the criteria for a diagnosis of MDS (“cytopenic control”); 3) 99 lymphoma-staging BMBx that were negative for hematolymphoid disease (“negative control”); 4) 95 MDS cases from MD Anderson Cancer Center for external validation.

Whole slide images (WSI) were tiled (224x224 pixels). Tiled images underwent a stain normalization process to standardize color variations. Each tile was then processed through a self-supervised learning framework, which has proven to be effective in capturing complex histomorphological features. The self-supervised pipeline effectively encoded complex morphological features into a structured, high-dimensional vector space. The high-dimensional latent space representations were subjected to dimensionality reduction using Uniform Manifold Approximation and Projection. Histomorphological phenotype clusters (HPCs) were then identified using the Leiden community detection algorithm with an unsupervised resolution selection, to detect meaningful clusters based on histomorphologic features. For each patient, the composition of these HPCs was calculated, resulting in a vectorized representation that reflected the frequency and distribution of the clusters within their tissue samples. These cluster compositions were then used to train a logistic regression model with 5-fold cross-validation, aimed at predicting the disease group.

Results:

The self-supervised model sorted H&E-stained image tiles into 31 clusters and reticulin-stain image tiles into 46 clusters. Evaluation of UMAPs showed that while MDS samples were enriched in certain clusters, samples from all categories were represented in each cluster and tiles from individual patient samples were represented in various clusters; these features suggest that BMBx from all groups exhibit some overlapping morphological characteristics. Enriched HPCs of H&E-stained BMBx in MDS cases captured histopathologically meaningful morphologic features including the presence of erythroid islands, dysplastic and clustered megakaryocytes, immature cells, increased vascularity, fibrosis, and cell streaming patterns in MDS cases. Conversely, depleted HPCs in MDS cases exhibited low cellular image tiles with lipid vacuoles and cells with normal morphology. A logistic regression model incorporating selected enriched and depleted H&E- and reticulin-stained BMBx in MDS showed an AUC of 0.92 in distinguishing MDS from negative controls, 0.92 in distinguishing MDS from cytopenic controls, and 0.83 in distinguishing MDS from other control groups. External validation using MDS cases from MD Anderson Cancer Center showed a similar pattern of HPC enrichment between the two cohorts. Evaluation of image tiles demonstrated the model's robustness to color variations between institutions. To determine the minimum tissue surface area required to achieve optimal diagnostic performance, we randomly selected varying numbers of tiles from each WSI. This analysis revealed that the minimum tissue area to achieve optimal performance for this model is approximately 0.8 mm².

Conclusion:

Self-supervised AI can distinguish BMBx from MDS patients from clinically relevant controls with high diagnostic performance. These promising results highlight AI's potential to enhance the accuracy and consistency of MDS diagnosis through the evaluation of bone marrow biopsies.

Loghavi:Pathology Education Partners; VJ HemeOnc, College of American Pathologists, OncLive, ICCS, MD Education, NCCN, MashUp Media, NCTN, Aptitude Health: Honoraria; Guidepoint; QualWorld; Gerson Lehrman Group, AlphaSight, Arima, Qiagen, Opinion Health: Consultancy; Astellas, Amgen: Research Funding; Abbvie: Current holder of stock options in a privately-held company; Syndx, Servier, BMS: Membership on an entity's Board of Directors or advisory committees; Abbvie, Daiichi Sankyo, BluePrint Medicine, Caris Diagnostics, Recordati, Servier: Consultancy. Park:Janssen Pharmaceutica NV: Other: Collection Cost associated with Material Transfer Agreement.