Development of a diagnostic technique for acute myeloid and promyelocytic leukemia using label-free 3D holo-tomographic image and deep learning

Abstract

Acute myeloid leukemia (AML) is a highly aggressive disease with unfavorable prognosis and a low survival rate. The current treatment method for AML is mainly chemotherapy, and drug efficacy depends on the type of leukemia. In particular, acute promyelocytic leukemia (APL) shows an excellent remission rate when treated with all-trans retinoic acid. Therefore, early diagnosis and suitable drug selection are the most crucial factors for the treatment. At present, medical teams preferentially diagnose leukemia from blood and bone marrow examination results. In this process, it lacks quantitative data for accurate classification. Several analytical techniques sequentially proceed for the final diagnosis, which are time-consuming. To overcome the shortcomings of conventional diagnosis procedures, we developed a novel diagnostic model based on holo-tomography and deep learning. From four AML and two APL cell lines, a total of 2,650 holo-tomographic images were obtained as input data. We trained and validated the classification model by applying “FishNet,” a deep learning architecture comprised of convolutional neural networks. The accuracy of the model was 94.7% for 6-class classification and 98.1% for binary classification. We expect that our model promptly provides reliable information from artificial intelligence in the stage of initial decisions for diagnosing AML.