VKCS: a pre-trained deep network with attention mechanism to diagnose acute lymphoblastic leukemia

Leukemia is a prominent hematologic malignancy that causes mortality and complications at different ages. In this paper, to provide a more accurate diagnosis of Acute Lymphoblastic Leukemia (ALL), a three-stage model based on transfer learning called variable-kernel channel-spatial attention (VKCS) is proposed. First, a deep pre-trained network extracts high-level features from the blood smear images. In the second stage, two attention mechanisms of variable-kernel spatial attention and variable-kernel channel attention consider spatial and channel information in parallel to improve model performance. The last step is the classification module. The experiments are repeated for different pre-trained networks, as well as for images in RGB, HSV, L*a*b*, and YCbCr color spaces, and by applying morphological operators (erosion and dilation) to images in different color spaces. The best model efficiency accuracy was achieved for images in HSV color space and the use of EfficientNet-V2M for feature extraction. The VKCS model accuracy is 100% for the ALL-IDB1 dataset and 99.6% for the ALL-IDB2 dataset, which are promising results.