EfficientNetB3-Based Transfer Learning Model for Accurate Classification of Acute Lymphoblastic Leukemia Blasts

Background: Acute lymphoblastic leukemia (ALL) predominantly affects pediatric patients and is characterized by the proliferation of immature lymphoblasts in the bone marrow. This uncontrolled growth impairs normal hematopoiesis, leading to anemia, immunodeficiency, and increased susceptibility to infections. Accurate detection and classification of these immature blasts are crucial for effective treatment planning and monitoring. Methods: This study utilizes transfer learning (TL) to improve the detection of immature ALL blasts in microscopic images. We employed the EfficientNetB3 model, a convolutional neural network (CNN) known for its efficient scaling and superior performance. The model was pre-trained on large datasets and fine-tuned with a dataset of 15,135 images from Kaggle, encompassing ALL-positive and ALL-negative samples. Image preprocessing techniques such as normalization, noise reduction, and segmentation were applied to enhance data quality. Results: The proposed TL model achieved a high training accuracy, indicating effective learning from the provided data. At epoch 19, the model's validation accuracy reached 97.75%, demonstrating strong generalization capabilities. The confusion matrix analysis showed high true positive and true negative rates, with minimal false positives and false negatives, underscoring the model's precision and recall. Conclusion: The use of TL with EfficientNetB3 significantly enhances the accuracy and reliability of detecting immature ALL blasts in microscopic images. This approach addresses the challenges posed by limited labeled data and image quality inconsistencies, providing a robust tool for improving leukemia diagnostics. The findings suggest that TL models can be instrumental in advancing clinical decision-making and patient outcomes in ALL treatment.