A reliable approach for identifying acute lymphoblastic leukemia in microscopic imaging

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Computational Biology

Background:

• Leukemia diagnosis is critical for patient survival but is labor-intensive.
• Automated systems using deep convolutional neural networks (CNNs) can assist in microscopic diagnosis.
• MobileNet, a pre-trained CNN, offers a compact and efficient solution for disease detection.

Purpose of the Study:

• To develop and evaluate an automated leukemia diagnosis system using a pre-trained MobileNet model.
• To enhance model robustness and prevent overfitting through advanced data augmentation and regularization techniques.
• To compare the proposed model's performance against existing methods on a public leukemia dataset.

Main Methods:

• Deployment of a pre-trained MobileNet (CNN) model for image classification.
• Implementation of L1 regularization and a novel dataset balancing approach (HSV color transformation, saturation elimination, Gaussian noise addition).
• Evaluation on the C_NMC_2019 dataset, including robustness testing with added Gaussian noise.

Main Results:

• The proposed MobileNet_M model achieved 95.33% accuracy and an F1 score of 0.95.
• The model demonstrated efficiency and robustness, performing well even with added Gaussian noise.
• Superior efficacy was observed compared to ALNet and other existing models on the same dataset.

Conclusions:

• The developed automated system provides an efficient and robust tool for leukemia diagnosis.
• The MobileNet_M model shows significant potential for supporting laboratory technicians and hematologists.
• The study highlights the effectiveness of advanced data augmentation and regularization in deep learning for medical imaging.