AI-Powered Evaluation of Bone Marrow Biopsies Can Distinguish Myeloid Neoplasms from Clinical Mimics

Background:

Histomorphological analysis of the bone marrow is an essential component in the diagnostic and classification framework for myeloid neoplasms, where the identification of dysplastic features and specific megakaryocytic characteristics are key criteria for establishing an accurate diagnosis. However, morphologic evaluation of bone marrow biopsies (BMBx) remains challenging and is characterized by significant inter- and intra-observer variability. We aimed to investigate whether AI-based image analysis of BMBx could enhance the accuracy of evaluating histopathologic features in myeloid neoplasms, potentially leading to more reliable diagnoses.

Design:

Bone marrow biopsies (BMBx) from patients with a clinicopathologic diagnosis of myelodysplastic syndrome (MDS) (n=90), myeloproliferative neoplasm [MPN, n=78, including PV (n=21), PMF (n=20), ET (n=16), CML (n=19)), non-MDS related cytopenia (n=55; “cytopenic controls"), and negative lymphoma staging BMBx (n=94, "negative controls") were selected and scanned at 40x magnification. Using the HALO AI platform, tissue and nuclear segmentation were performed, and an algorithm was trained to evaluate the following features: 1) cellularity and cell density at the tissue level; 2) megakaryocyte (MK) size, MK nuclear optical density, MK nuclear-to-cytoplasmic (N/C) ratio, MK nuclear roundness, and MK density per mm².

To capture more abstract and complex morphological features, image tiles were generated through AI-based annotation of megakaryocytes and nuclei, as identified by the trained tissue classifiers. Feature extraction was performed using a pre-trained convolutional neural network, and unsupervised clustering was applied to the vector representations of the image tiles to identify distinct histomorphologic features among megakaryocytes and nuclei. A Random Forest classifier with 5-fold cross-validation was then trained using the cluster composition of tiles for each patient to develop an ensemble predictive model that integrates supervised features with unsupervised features related to megakaryocytes and nuclei.

Result:

Cellularity was significantly higher in MDS cases compared to the control groups. Additionally, MKs in MDS cases were notably smaller, exhibited a higher density per mm², and had a greater number of nuclei per MK. Unsupervised clustering of all nuclei and MKs revealed 36 distinct clusters of MKs, and 44 distinct clusters of nuclei based on their histomorphological features. Expert visual assessment of the image tiles confirmed that the model accurately captured meaningful morphological features including alterations in MK size, shape, and nuclei compatible with dysplasia. By integrating both supervised and unsupervised features in an ensemble model, ROC curve analysis showed mean AUC values of 0.89, 0.91, and 0.86 for distinguishing MDS from all other categories, negative controls, and cytopenic controls, respectively. The model achieved an AUC of 0.92 for distinguishing MPN cases from the negative control group and AUC of 0.86 for distinguishing MPN from MDS cases.

Conclusion:

We employed AI-based image analysis to determine whether BMBx histologic features can accurately identify characteristics associated with myeloid neoplasms, and effectively differentiate these cases from clinical mimics. Our model, leveraging AI-based quantified features showed an AUC of 0.91 and 0.93 in distinguishing MDS and MPN cases from negative controls, respectively. Additionally, the model demonstrated an AUC of 0.86 in distinguishing MDS from cytopenic controls, highlighting its effectiveness in challenging diagnostic scenarios. These findings suggest that AI-based image analysis of BMBx has the potential to reduce subjectivity and improving consistency in the diagnosis of myeloid neoplasms.