Detection of Acute Lymphoblastic Leukemia Using CollateNet

Acute Lymphoblastic Leukemia (ALL), a malignant blood cancer is very common in children. This cancer shows good response to treatment provided it is diagnosed in time. As this is a fast-growing cancer, timely diagnosis of this cancer is crucial. The diagnosis of this cancer requires identification of unhealthy lymph blast cells. As the unhealthy lymphoid blast cells are morphologically similar to the healthy lymphoblast cells, distinguishing between them is a very difficult task and requires advanced methods to be applied. Deep Learning algorithms like Convolution Neural Network (CNN) have shown great results in image classification tasks but they are prone to overfitting. Bearing this in mind, this paper proposes CollateNet, a fully convolutional network which uses collation blocks. This architecture has been applied on CNMC 2019 dataset to classify single cell blood smear images into ALL and normal cells. The proposed architecture achieved 89.88 % and 87.96% training and validation accuracies respectively and a training and validation Precision of 92.58% and 92.24% respectively. The model is able to overcome the challenge of overfitting and vanishing gradient due to the collation blocks used in the neural network. The post training and post testing analysis shows the effectiveness of the proposed architecture for timely detection of ALL from single cell blood smear images and hence can be used in future for the diagnosis of ALL at an earlier phase.