<p>Leukemia (blood cancer) is one of the diseases with a high mortality rate worldwide. It occurs in various kinds of blood cells but most often starts in White Blood Cells (WBCs). Acute Lymphocytic Leukemia (ALL) is one of the four main kinds of leukemia that grow from primary (immature) forms of lymphocytes and happen in all ages. ALL will lead to death within a few months if it is not detected and treated quickly. Hematologists analyze blood samples under a microscope to identify leukemia in the Manual (traditional) method. This procedure is slow, time-consuming, less accurate, and dependent on hematologists’ experts. Computer-assisted systems can help pathologists to overcome these challenges. In this article, we proposed two kinds of Ventral-Dorsal Attention Blocks (VDAB) for automated classification of ALL based on the Two Visual Streams Hypothesis (TVSH). First, the Parameterized Ventral-Dorsal Attention Block (PVDAB) is designed that composed of the Parameterized Ventral Attention Block (PVAB) and the Parameterized Dorsal Attention Block (PDAB). The extracting channels related to the shape of WBCs and focusing on their location in the selected channels are performed by the PVAB and PDAB, respectively. Then, the PVDAB is improved and the first light residual cognitive attention block, Non-parameterized Ventral-Dorsal Attention Block (NPVDAB), is introduced that doesn’t impose any learning parameters on the networks. The PVDAB and NPVDAB can be employed in the architecture of each Convolutional Neural Network (CNN). We embedded the suggested attention blocks in the architectures of the ResNet18 and MobileNetv1, and four attention-based networks were generated. Different data augmentation techniques are performed on the ALL-IDB2 dataset to avoid model overfitting and its generalization. The fine-tuned ResNet18, MobileNetv1, and attention-based networks are trained, validated, and evaluated on the dataset with similarity parameters for 40 epochs. The experiment results indicated that the network-3 (ResNet18+NPVDAB) achieved better performance metrics than others with accuracy and an F1-score of 99.33% in the test step.</p> Graphical abstract <p></p>