FResFormer: Leukemia Detection Using Fusion-Enabled CNN and Attention

In this paper, the binary classification of normal and abnormal (malignant) cells from the microscopic images is done. It is quite challenging due to the appearance of both cells morphologically similar. In particular, manual identification of malignant and benign cells from the microscopic images in early stages is difficult because of the similar resemblance of both cells in their appearance. This early diagnosis process requires advanced techniques like flow cytometry that are currently used and more expensive and, therefore, are not accessible in all places. Additionally, a medical expert is also required. Therefore, by using automated diagnostic tools, we can perform better diagnoses with low-cost microscopic image data. In this paper, we propose a classification of normal and malignant cells in the lymphocytes using a fusion-enabled CNN and Attention-based neural network, which we named Fusion-based Residual Transformer (FRESFORMER) architecture. Our proposed model tops the performance on the benchmark ISBI 2019 challenge dataset. The proposed model achieves an F1-Score of 84.89.