A novel Deep Learning Framework (DLF) for classification of Acute Lymphoblastic Leukemia

Cancer of blood, more specifically known as Leukemia, is a deadly disease that is responsible for the abnormal proliferation of immature white blood cells in bone marrow. Due to sudden surge in the speed of disease progression and lack of timely diagnosis, the chance of disease remission also diminish. Leukemia incidence rate has seen a steep rise of 144.7% in past 22 years from 28,700 in 1998 to 60,530 in 2020 (Hao et al. Sci Rep 9(1):1–13, 2019). Deep learning has revolutionised the solution to classification problems, more specifically, image based classification. Thus, in this paper, we propose a novel deep learning framework(DLF) based on convolution neural network for the diagnosis of Acute Lymphoblastic Leukemia (ALL), which is one of the four types of leukemia. The proposed method does not require feature extraction.Also it does not require any pre-training on any other database and thus can be used for real time application for detection of leukemia. The proposed framework is simple compared to existing deep networks. Number of free parameters to be tuned in the proposed framework is 41,626, which is quite less than the learnable parameters in the existing pre-trained complex networks such as, AlexNet (over 60 millions), Visual Geometry Group(VGG-Net)(138 million), Residual Network(ResNet 152)(60.3 millions). As the number of free parameters of the proposed framework is approximately 1400 times less than those of existing deep networks, the simulation of the proposed framework has been possible on simple processor(i5 @2.53GHz processor with 4 GB RAM), it does not require any GPU for processing. Despite of lesser number of free parameters in the proposed model, it is able to diagnose the disease with 100% accuracy in most of the repetitions and an average accuracy of 98.17% .This is the average of two other averages i.e., Experiment-A (98.62%), obtained when the data has been partitioned on (80%-20%) training and testing sets for 20 epochs and Experiment-B(97.73%) when the data has been partitioned as 60-48 training and testing images for 30 epochs.