Timely Diagnosis of Acute Lymphoblastic Leukemia Using Artificial Intelligence-Oriented Deep Learning Methods.

Area of Science:

• Computational intelligence
• Medical image analysis
• Machine learning in oncology

Background:

• Leukemia is a fatal cancer with subtypes like acute lymphoblastic leukemia (ALL).
• Early ALL diagnosis significantly impacts treatment outcomes.
• Computational intelligence offers rapid identification and classification of ALL.

Purpose of the Study:

• To develop and evaluate a novel convolutional neural network (CNN) for classifying acute lymphoblastic leukemia (ALL).
• To compare the performance of the proposed CNN against established deep learning models (ResNet-50, VGG-16) and traditional machine learning techniques for ALL detection.

Main Methods:

• Utilized a dataset from a CodaLab competition for classifying leukemic cells in microscopic images.
• Employed and retrained ResNet-50 and VGG-16 deep learning networks.
• Proposed a custom 10-layer convolutional network with max-pooling layers.
• Developed six common machine learning techniques for binary classification of ALL.

Main Results:

• The proposed CNN achieved a validation accuracy of 82.10%.
• VGG-16 and ResNet-50 achieved validation accuracies of 84.62% and 81.63%, respectively.
• Among machine learning methods, Random Forest yielded the highest accuracy at 81.72%, while Multilayer Perceptron had the lowest at 27.33%.

Conclusions:

• The proposed convolutional neural network demonstrates optimal accuracy and efficiency for ALL diagnosis.
• The CNN offers a less complex alternative to pre-trained networks, suitable for clinical integration.
• This approach provides good performance and optimal execution time for leukemia diagnosis in clinical systems.