Acute lymphoblastic leukemia Detection and classification of its Subtypes using pre- trained deep Convolutional neural network

Abstract

Leukemia is a form of blood cancer which affects the white blood cells and damages the bone marrow. Usually complete blood count and bone marrow aspiration is used to diagnose the acute lymphoblastic leukemia. It can be a fatal disease if not diagnosed at the earlier stage. In practice, manual microscopic evaluation of stained sample slide is used for diagnosis of Leukemia. But, manual diagnostic methods are time consuming, less accurate and prone to errors due to various human factors like stress, fatigue etc. Therefore an efficient and robust computeraided method for diagnosing of acute lymphoblastic leukemia is needed to be developed. We deployed deep convolutional neural network for automated detection of acute lymphoblastic leukemia using microscopic blood images. In contrary to the training from scratch, we deployed pre-trained AlexNet which was fine-tuned on our datasct. Last layers of the pre-trained network were replaced with new layers which can classify the input images into normal or cancerous. To reduce overtraining, data augmentation technique was used. We also compared the datasets with different color models to check the performance over different color images. The average accuracy for ALL detection was 99.50%. The subtype classification of Acute Lymphoblastic Leukemia based on French-American-Brit classification is necessary for prognostication and treatment of acute lymphoblastic leukemia. For classification of its subtypes deep convolutional neural network was trained on the dataset which classify the input images into 4 classes i-e L1, L2, L3 and Normal, which was mostly neglected in previous literature. For ALL subtypes classification we were able to achieve the accuracy of 96.06%. Each method is studied separately and different experiments are conducted to evaluate their performances. 'Ne have compared our proposed method with previous literature in term of accuracy and number of images in their dataset. Unlike the standard methods, our proposed method was able to achieve high accuracy without any need of microscopic image segmentation.