2024 Volume 73 Issue 1 Pages 69-77
Deep learning is a machine learning technique using a multilayered neural network that mimics the neural circuits in the human brain. Deep learning models are capable of automatic identification of hidden features in data sets following numerous training iterations, ranging from dozens to hundreds, that result in the generation of an accurate and efficient decision model. The immature granulocyte or blast cell population in peripheral blood is a critical indicator for differentiating hematopoietic diseases. Thus, a highly accurate automated screening technology needs to be established. This study examined the training conditions in generating immature granulocyte recognition artificial intelligence models using convolutional neural networks. We examined the clinical potential of the generated artificial intelligence models. This study used five types of ResNet models with a layer count ranging from 18 to 152 layers. Transfer learning and fine-tuning were performed using 6727 labeled blood cell images and eight types of optimizers to generate immature granulocyte recognition artificial intelligence models with optimal weighting. Clinical assessments were performed on 25 healthy and 25 cases with the appearance of immature granulocytes using the trained artificial intelligence models. The minimum–maximum total accuracy ranges were 0.9131–0.9788 for the healthy cases and 0.8177–0.8812 for the immature granulocyte cases. Our convolutional neural network-based immature granulocyte recognition artificial intelligence model had an accuracy rate of 97% for healthy cases and a rate higher than 88% for cases with the appearance of immature granulocytes. These findings indicate that our artificial intelligence model is a useful morphological differentiation support technology for peripheral blood smear screening.
