AML leukocyte classification method for small samples based on ACGAN.

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Acute myeloid leukemia (AML) diagnosis relies on microscopic analysis of peripheral blood smears.
• Accurate identification and counting of white blood cells are crucial for diagnosing hematologic malignancies like AML.
• Deep learning (DL) methods for blood cell recognition require substantial datasets, posing challenges for small sample sizes.

Purpose of the Study:

• To evaluate the effectiveness of Auxiliary Classification Generative Adversarial Network (ACGAN) for white blood cell classification using limited data.
• To compare ACGAN's performance against traditional classifiers and current state-of-the-art methods in small-sample scenarios.

Main Methods:

• ACGAN was trained on the TCIA dataset for white blood cell classification.
• Performance was assessed using accuracy, precision, recall, and F1 score.
• Results were benchmarked against two classical classifiers and advanced deep learning techniques.

Main Results:

• ACGAN achieved 97.1% accuracy on the validation set.
• Precision, recall, and F1 scores for ACGAN were 97.5%, 97.3%, and 97.4%, respectively.
• ACGAN demonstrated superior or competitive classification accuracy compared to other advanced methods.

Conclusions:

• ACGAN is a viable and competitive deep learning method for classifying small white blood cell sample datasets.
• The approach shows significant potential for improving the accuracy and efficiency of AML screening and diagnosis.
• ACGAN addresses the data limitations often encountered in applying deep learning to medical image analysis.