​簡介

Over the past decade, advancements in artificial intelligence (AI) and medical technologies have transformed clinical practices, particularly in hematopathology. Traditional blood smear microscopy counts have largely been replaced by automated digital blood cell analyzers, which have greatly reduced the workload for medical personnel. Despite these advances, the precise identification of abnormal cells still depends on the expertise of seasoned morphologists. In a groundbreaking study, we harnessed a dataset of 47,046 images of six common white blood cell types obtained from an automated digital blood cell image analyzer to train a deep learning model. The model's performance was then tested against two separate sets of images and compared with the assessments made by clinical examiners. The results were impressive, with the machine learning model achieving an accuracy rate of over 0.98 for cell identification. Furthermore, when the model was adapted for smartphone use via transfer learning to a Mobile Vit small model, it maintained a robust performance with an accuracy above 0.87. Of particular note is the model's capability to identify Blast cells—an abnormal cell type—with a high degree of precision, reaching an accuracy of 0.93. This not only demonstrates the model's potential to assist less experienced examiners, but it also suggests that it can match the proficiency of morphologists with up to four years of experience. In fact, examiners with less than two years of experience have an accuracy rate of less than 0.7 in identifying Blast cells, underscoring the significant edge provided by the machine learning model. In conclusion, this study indicates that AI-enhanced digital analyzers can not only supplement the preliminary identification tasks traditionally carried out by automated analyzers but can also achieve recognition accuracy comparable to that of highly experienced professionals. This could herald a new era in clinical examination, where AI plays a crucial role in ensuring accuracy and efficiency in blood cell morphology analysis.  

Keywords: Deep learning, blood cell morphology, AI, Smart phone.