Acute lymphoblastic leukemia (ALL) diagnosis is a challenge, including invasive classical methods, which are time-consuming, inaccurate, and error-prone. In this paper, we propose modifying the VGG16 architecture to improve its performance in the classification task. We utilize the Acute Lymphoblastic Leukemia (ALL) image dataset to train the proposed modified VGG16 model. The dataset was split into training and testing sets at a ratio of 80% for training data, 10% for validation data, and 10% for testing data. The ALL dataset consists of four classes: Benign, Early, Pre, and Pro. The results of the proposed modified VGG16 model were very satisfactory: accuracy, 96.59%; precision, 96.61%; sensitivity, 96.59%; F1-score, 96.58%; and Matthew's correlation coefficient of 95.35%. It was demonstrated that the image size also influences the model, indicating a trade-off based on how efficient a computational one can be concerning classification accuracy. These findings highlight the promise of deep learning algorithms to revolutionize all characterizations and offer potential utility for future applications in medical imaging.