AI-Based Detection of Myelodysplastic Syndrome from Neutrophil Morphology in Peripheral Blood

Introduction

Myelodysplastic syndrome (MDS) is a potentially fatal bone marrow stem cell disorder leading to cytopenias and risk of progression into acute myeloid leukemia (AML). Diagnosis of MDS is challenging and requires multi-modal, iterative procedures, such as bone marrow cytology, molecular genetics and cytogenetics. Currently, no rapid point-of-care tests are available in routine clinical practice. Computer-vision based detection of specific morphological cues in peripheral blood cells have provided promising results and a rationale for systematically evaluating machine-learning algorithms to accelerate MDS diagnostics. Here, we aim at developing a computer vision pipeline to diagnose MDS by analyzing single-cell images taken routinely from standard hematology cell analyzers from peripheral blood smears.

Methods

Digitized cell images of peripheral blood smears from 60 healthy individuals and 60 MDS patients were classified (CellaVision Software), producing 77'088 single-cell image patches with cell type annotations. Diagnosis of MDS patients had been previously confirmed by bone marrow biopsy, molecular and cytogenetics according to WHO guidelines. Of these, 45 patients exhibited non-prominent MDS (npMDS), meaning experts saw no overt dysplasia in the cell morphology in peripheral blood neutrophils. Neutrophils from the remaining prominent MDS (pMDS) samples were expert-annotated as dysplastic or benign for each MDS patient. The classifier was developed on 8'478 annotated neutrophils from 24 healthy individuals and 14 patients with pMDS. Of these, pMDS patients were split amongst the train, validation, and test sets at a ratio of 5:5:4, and healthy individuals at a ratio of 11:6:7. Neutrophil counts in the train, validation, and test sets were 5'804, 1'823, and 851 cells, respectively. A cell segmentation algorithm was trained to identify the central neutrophil in each image patch. Synthetic backgrounds were generated to exclude peripheral cells, promoting the prediction of dysplasia from sub-cellular features. Finally, a deep neural network detected dysplasia at the neutrophil level, and single-cell scores were aggregated using a naive threshold.

Results

We developed a neutrophil classifier to diagnose MDS by detecting morphological features related to dysplasia. Our model detected dysplastic neutrophils with a sensitivity and specificity of 0.95 and 0.89, respectively. Of note, removing the 44 cells from pMDS samples annotated as benign but classified by the model as dysplastic increased the specificity to >99%. Upon aggregating cell-level predictions to the sample level, MDS status was correctly predicted for every sample in the validation and test sets. The classifier was then applied to a previously unseen dataset of 26,164 un-annotated neutrophils from 44 healthy and 43 non-prominent MDS patients, i.e., morphologically non-evaluable MDS, to predict MDS diagnosis. The model identified MDS with a sensitivity and specificity of 0.86 and 1.0, respectively. Attention-maps of analyzed cells identified nuclear texture, shape and granularity in the cytoplasm as essential markers for model performance.

Conclusion

MDS could be predicted with high sensitivity and specificity in patients with both morphologically visible and non-prominent dysplasia by analyzing the morphology of peripheral blood neutrophils with computer vision. Our model represents a step forward in rapid detection of MDS based on cell morphologies in peripheral blood smears to accelerate and streamline MDS diagnostics in clinical practice. To improve performance and enhance explainability, larger datasets, including other hematological disorders, are needed. To define the source of the failure (model classification vs. human expert annotation), failure modes of incorrectly classified cells are currently being characterized.

Romano:Moonlight AI: Current equity holder in private company. Habringer:Pentixapharm: Consultancy, Other: Safety Review Committee Chairman; Moonlight AI: Current equity holder in private company; Incyte: Honoraria, Other: Advisory Board Ponatinib. Ruiz:Moonlight AI: Current equity holder in private company.