Vision Transformer Features-Based Leukemia Classification

Acute Lymphoblastic Leukemia (ALL) is a disease that is caused by the uncontrollable growth of immature and malignant White Blood Cells (WBCs) which are called lymphoblasts. It occurs when the bone marrow contains 20% or more lymphoblasts. Therefore, Leukemia is diagnosed by counting White Blood Cells (WBCs) in the microscopic smears of bone marrow and blood. There are several attempts for effective leukemia classification with the help of computer aided-based methods by analyzing microscopic smear images. However, how to further improve the classification accuracy is still challenging. Therefore, in this paper, we propose an end-to-end framework to classify ALL. First, we train a more robust specific pre-trained Vision Transformer (ViT) trained on the same type of data, ALL-IDB1. Then, we examine the extracted features of Deep ViT across layers and utilize these features as dense descriptors to make predictions. Based on the observation in this paper, we find and empirically demonstrate that such features, when extracted from the ViT model, consistently improve the classification performance. Moreover, we observe that the pre-trained ViT which is fine-tuned on specific dataset is more effective than the general images dataset. We show that our proposed method achieves an average accuracy of 98.75% which is a competitive result with recent state-of-the-art by a large margin.