Acute myeloid leukemia diagnosis using deep learning
QamerT Al-Rashedi, Eman M. Nagiub, Khaled F. Hussain, Nagwa M. Omar · The Egyptian Journal of Haematology · 2020
Background Early disease detection has a great impact on saving lives. One of the top deadliest diseases in the world is leukemia. Once detected, its treatment is immediately required. Standard morphologic diagnosis of leukemia by hematologists is done by examining the patient’s peripheral blood (PB) and bone marrow (BM) under microscope. However, manual recognition is prone to variations such as experience and tiredness where the percentage of error in diagnosis is between 30%–40%. Therefore, it is necessary to have a robust automated system for leukemia detection that is not influenced by human variations. Machine learning is attracting interest in the biomedical field as it improves sensitivity and specificity of the disease diagnosis, influencing the objectivity of the decision-making process.Objectives Testing deep learning technique that use microscopic images to identify leukemia based on a pretrained deep convolutional neural network which is an approach of machine learning algorithms.Material and method Dataset used was collected from Clinical Pathology lab, Assiut University Hospital, Egypt, containing binary balanced dataset: leukemic and normal blood, each sample composed of PB and BM smears images. We analyzed 412 digital images; 206 images of leukemia (AML types only) & 206 images of normal blood. 1-Training & testing phase: testing different types of pre-trained convolutional neural networks models ex. Alexnet, VGG16, GoogleNet, ResNet101, and Inception-v3. We divided images into two groups; 80% of images assigned as “training-set”, each of the models trained on the images to extract features to differentiate leukemic and normal images, remaining 20% images were assigned as “testing-set” or so called “never seen” by the models. 2-Evaluation of sensitivity/specificity/accuracy: evaluating ability of the models to detect target; based on sensitivity, specificity and accuracy of detection. 3. We computed the testing execution time per image in dataset.Results Comparing the ability of models to differentiate leukemia vs normal images showed that Inception-v3 model had the highest accuracy (99%) in detection and classification of AML. In terms of sensitivity and specificity, the Inception-v3 can detect all leukemic cases of AML, with 100% sensitivity; and specificity of 97.8%. Inception-v3 required only 0.2273 seconds to test each image in AML-IDB.Conclusion Inception-v3 outperforms other pretrained convolutional neural networks in diagnosis of leukemia, this model of Machine learning algorithm can be used in the context of lab diagnosis acting as a second opinion after manual evaluation of leukemia.