Toward automated classification of b-acute lymphoblastic leukemia

Abstract

We present an automated approach of classifying a stain-normalized white blood cell as being either a malignant (B-ALL) cell or a healthy (HEM) cell. We apply various data augmentations to the images in order to compensate for the class imbalance. As our primary approach, we use the ResNeXt101 Convolutional Neural Network architecture and “cut” it at various points in the network. We show that an ensemble of the same ResNeXt101 model trained for varying durations and a change in the decision threshold boundary lead to better performance as compared to individual model performances. Our final ensemble model leads to a weighted F1 score of 0.825 on the online preliminary test set. After training on the entire training data and validating in the preliminary test set, ResNeXt50 and ResNeXt101 achieve weighted F1 scores of 0.857 and 0.849 on the preliminary test set, respectively.