DeepMEN: Multi-model ensemble network for b-lymphoblast cell classification

Abstract

In recent years, convolutional neural networks have achieved great results in many image classification tasks, including the field of medical image processing. However, for leukemic B-lymphoblast cells, the traditional convolutional neural network has some problems. On the one hand, there is a strong individual difference in the cell image, that is, there is a big difference between the cells of different subjects, and this difference may exceed the difference between abnormal cells and normal cells. On the other hand, the two cell types appear very similar. People without expertise can’t tell if a cell is abnormal. In order to solve the above problems, we proposed Deep Multi-model Ensemble Network (DeepMEN). We trained a series of deep learning models and selected the six models with the most potential as candidates. We trained the six models respectively and used the model ensemble technology to fuse the six models. In addition, we use the pseudo-label and Test Time Augmentator (TTA) to reduce covariance shifts caused by individual differences. Finally, we obtained a weighted F1-score of 0.903 in the preliminary test set and 0.8856 in the final test set.