Classification of leukemic B-Lymphoblast cells from blood smear microscopic images with an attention-based deep learning method and advanced augmentation techniques

Abstract

Acute Lymphoblastic Leukemia (ALL) is a blood cell cancer type characterized by an increased number of immature lymphocytes. It is one of the leading cancer inflicted causes of death among children and affected around 876.000 people globally in 2015. In this work, we describe methods used for the “Classification of Normal versus Malignant Cells in B-ALL White Blood Cancer Microscopic Images” ISBI challenge. Due to the morphological similarity between malignant and normal cells, the classification of these cell types is a very challenging task, even for human experts. Deep learning using Convolutional Neural Networks (CNN) provides state-of-the-art techniques for image classification, even with limited annotation data, based on techniques like transfer learning and data augmentation. Our solution to tackle this classification problem is based upon advanced augmentation techniques to counter overfitting the small dataset and transfer learning with adaptive learning rates. Additionally, we incorporated a basic attention mechanism based on a region proposal subnetwork to boost our results for this competition. The ResNet 18 network with an additional regression head achieved a weighted F1 score of 0.8284 on the final test set and 0.8746 on the preliminary test set.