[2J1-GS-10a] AI application

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)

[2J1-GS-10a-03] Evaluation of Bone Marrow Blood Cell Detection Using YOLO with Deep Learning

〇Takao Naito¹, Yuki Horiuchi¹, Maki Sugiura¹, Akihiko Matsuzaki¹, Naoki Kosaka², Yoko Tabe¹ (1. Graduate School of Medicine, Juntendo university, 2. School of Medicine, Juntendo university)

Keywords:

Bone Marrow Hematopoietic Cell,Deep Learning,Object Detection,YOLO,Image Processing

In bone marrow examinations, which are essential for the diagnosis of hematological disorders including hematopoietic malignancies, morphological classification of hematopoietic cells on bone marrow smear specimens is routinely performed. To support this process using machine learning–based cell classification, it is necessary to extract images of individual hematopoietic cells from smear specimens. However, in bone marrow smears, hematopoietic cells are often densely clustered and intricately attached to each other, making automatic detection of individual cells from microscopic images technically challenging. In this study, we aimed to develop an algorithm that automatically detects hematopoietic cells on smear specimens and separates tightly clustered cells. From whole-slide images of the specimens, we extracted all hematopoietic cells, whether isolated or adherent, and attempted individual separation of clustered cells using the YOLO object detection model. The performance of the YOLO v9c model trained on 1,199 cell images achieved a sensitivity of 0.760 and a precision(mAP@0.5) of 0.897. Furthermore, visual evaluation using eight bone marrow smear images demonstrated an overall cell detection rate of 94.0% and a separation rate of 75.6% for clustered cells.