I-Net: a deep CNN model for white blood cancer segmentation and classification

The immune system relies on white blood cells and platelets, which are both produced in the bone marrow and together account for around one percent of the blood corpuscles. Acute lymphoblastic leukaemia (ALL) and acute myeloid leukaemia (AML) are two major subtypes of acute leukaemia identifiable from its lineage. Unlike other chronic diseases, leukaemia is a curable blood disorder and patients' survival is possible with precise treatment. The effectiveness of this disease's treatment can be greatly influenced by early diagnosis. This study focused on a deep neural network for the segmentation and classification of ALL using the SN-AM and ALL-IDB datasets and obtained from the cancer imaging archive (TCIA) repository. ResNet-50 and VGG-19, two of the most popular deep learning networks, were used. The use of stored weights was not used for these two networks; instead, we modified the weights and learning parameters. A UNet with InceptionV2 model was used for the segmentation, while convolutional neural network (CNN) was employed to train the images after feature selection. An improved CNN called i-Net with more convolutional layers and tuned hyperparameters was proposed for the classification into normal and cancerous white blood cells. Data augmentation, dropout regularization, and batch normalization were employed to reduce overfitting. ResNet-50, VGG-19, and a proposed deep neural network called “i-Net”, all have validation accuracy of 92.2%, 92.3%, and 99.18%, respectively. However, when trained without early stoppage, the model (i-Net) accuracy decreased after the 30th training cycle (epoch). CNN has shown to be accurate at diagnosing ALL according to this study. When compared to other pre-trained deep learning models such as the standard VGG-19 and ResNet-50, we achieved a better performance on the test dataset of about 630 microscopic images suggesting that the CNN can be used in clinical decision support systems (CDSS) for leukaemia detection.