A machine-learning-based algorithm for bone marrow cell differential counting

Abstract

Background

Differential counting (DC) of different cell types in bone marrow (BM) aspiration smears is crucial for diagnosing hematological diseases. However, a clinically applicable method for automatic DC has yet to be developed.

Objective

This study developed and validated an artificial intelligence (AI)-based algorithm for identifying and classifying nucleated cells in BM smears.

Methods

In the development phase, a mask region–based convolutional neural network (Mask R-CNN)-based AI model was trained to detect and classify individual BM cells. We used a large data set of expert-annotated images representing a variety of disease categories. The BM slides were stained with Liu’s stain or Wright–Giemsa stain. Consensus meetings were held to ensure experts from different institutes applied consistent criteria in classifying cells. Subsequently, the performance of the AI algorithm in identifying cell images and determining cell ratios was evaluated using a multinational clinical dataset.

Results

The AI model was trained on 542 slides (85.1 % stained with Liu’s stain and 14.9 % with Wright–Giemsa stain) containing 597,222 annotated cells. It achieved an accuracy of 0.94 for the testing dataset containing 26,170 cells. The performance of the AI model was further validated using another multinational real-world dataset (data obtained from three centers in Taiwan and one in the United States) comprising 200,639 cells. The AI model achieved an accuracy of 0.881 in classifying individual cells and demonstrated high precision in classifying blasts (0.927), bands and polymorphonuclear neutrophils (0.955), plasma cells (0.930), and lymphocytes (0.789). When the differential counting percentage of each cell type was assessed, a strong correlation (ρ > 0.8) between the AI and manual methods was observed for most cell categories.

Conclusions

In this study, an AI algorithm was developed and clinically validated using large, multinational datasets. Our algorithm can locate and classify BM cells simultaneously and has potential clinical applicability for automating BM differential counting.