AI-Powered Auto-Classification of Blood Cells in Bone Marrow Smear Testing
Hiroyuki Nozaka, Haruto Okuse, Aono Nakata, Mihoko Kushibiki, Shu Ogasawara, Masahiro Ishiyama, Kosuke Kamata, Kazufumi Yamagata · Studies in health technology and informatics · 2025
Recent advancements in AI technology present significant potential for improving the accuracy and efficiency of hematological testing. This study developed an AI-based classification model for blood cells in bone marrow smears, utilizing ResNet and ResNeXt architectures. The dataset consisted of 7376 images of nucleated blood cells from 46 healthy individuals, manually classified into 14 categories by experts. The ResNeXt-101 model, which was the most accurate in clinical assessments, achieved a total accuracy of 83.57%, with frequent misclassifications occurring between closely related cell categories. The transitional stages between immature and mature cells present a challenge for classification, suggesting the need for a multistage AI model to improve accuracy, although first screening with this AI model may be useful in enhancing testing efficiency.