The study presents the deep learning-based XIncept-ALL model for automatic diagnosis and severity classification of acute lymphoblastic leukemia (ALL), which mainly affects white blood cells in children. The model combines the InceptionV3 and Xception networks through feature fusion blocks and uses automatic data expansion to address class imbalance and avoid overlearning. Grad-CAM visualizations show that the model targets clinically relevant areas of cells. The XGBoost classifier divides cells into four categories: Benign, Early, Pre and Pro. They tested the model on a new Pak-ALL file from Pakistani hospitals and other sources, including an Iranian external file. It achieved an average accuracy of 99.5% on this challenging external data set. The results confirm the effectiveness of the model for clinical decision support.