Background:Acute myeloid leukemia (AML) is a prevalent hematological malignancy worldwide. The estimated incidence of AML was 2.4 per 100,000 standard population in Taiwan in 2016. Bone marrow studies are essential for the initial workup in most cases. However, these examinations require a thorough evaluation of the histopathology slides by hematopathologists. Acute promyelocytic leukemia (APL) is a subtype of AML. APL has completely different initial treatment and prognosis from the other subtypes of AML. Convolutional neural network (CNN) is the state‐of‐the‐art technology of image recognition and has been shown to successfully predict common gene mutations by using cancer pathology images (Nicolas et al. 2018 Nature Medicine).Aims:We planned to use the whole‐slide scanning images of bone marrow aspiration smears to examine the utility of CNN in predicting APL in AML patients.Methods:We collected the whole‐slide scanning images of patients with newly diagnosed AML at Taipei Veterans General Hospital between November 2009 and September 2017. The final cohort consisted of 261 AML cases. The patients’ age, sex, bone marrow blast percentage, cytogenetics and molecular studies, such as FLT3‐ITD and NPM1 and outcomes, were collected. The cohort was split into a training set (80%) and a testing set (20%). In addition, two AML cohorts from Far Eastern Memorial Hospital (n = 19) and Kaohsiung Veterans General Hospital (n = 29) were served as external validation cohorts. Inception V3 was used as the baseline model. Stochastic gradient descent was employed to refine the weights in the CNNs using the training set. The finalized model was tested on the held‐out test set. Prediction accuracy and the area under the receiver operating characteristic curves (AUCs) were analyzed to evaluate the performance of the algorithms. All CNN models were trained using Keras version 2.1.5 with Tensorflow 1.7.0, and statistical analysis was performed with R version 3.5.1.Results:The median age of the AML patients at diagnosis was 65 (range 17–96) years and 58.9% of them were male. CNN classified APL from the other subtypes of AML with high accuracy (AUC 0.92). In addition, CNN also could predict other chromosomal abnormalities, such as complex karyotypes, and molecular abnormalities, such as NPM1 mutation. Furthermore, CNN successfully predicted patients’ overall survival by using its classification.Summary/Conclusion:CNN successfully predicted genetic alterations of AML patients by using the histopathology images of the bone marrow aspiration smears. These results demonstrated the utility of convolutional neural networks in predicting common genetic mutations. Our developed methods are generalizable to other hematological malignancies or other cancers.image
Read full abstract