This study investigates a new approach to improving diagnosis for ALL using Convolutional Neural Networks. In this work, a hybrid approach was proposed, merging the Xception and DenseNet121 architecture to classify images of blood cells. The approach should try to address critical problems that include limited Labeled data, class imbalance, and variability in image quality that might impede precise detection. The hybrid CNN model itself automatically extracts the critical features, thereby considerably improving the diagnostic accuracy. The model was tested on a dataset of blood smear images that resulted in 99.69% training accuracy and more than 97% validation accuracy. These results emphasize the strength and efficiency of this model in distinguishing ALL-positive samples from ALL-negative ones. The research presented here extends the knowledge of automated diagnostic tools by proposing a more effective and consistent method for early and accurate diagnosis of ALL and provides the foundation for further developments in diagnostics related to leukemia.