Quantifying imbalanced classification methods for leukemia detection.

Area of Science:

• Hematology
• Computational Biology
• Machine Learning

Background:

• Acute Lymphoblastic Leukemia (ALL) is characterized by uncontrolled B-lymphoblast proliferation.
• Early detection of malignant B-lymphoblasts in bone marrow is crucial for effective treatment.
• Automated cell classification is difficult due to fine-grained variability and imbalanced data.

Purpose of the Study:

• To explore State-Of-The-Art (SOTA) deep learning approaches for classifying imbalanced leukemia datasets.
• To evaluate input-based, Generative Adversarial Network (GAN)-based, and loss-based methods for leukemia detection.
• To identify the most effective deep learning strategy for mitigating class imbalance in leukemia classification.

Main Methods:

• Utilized deep learning algorithms to address fine-grained classification challenges in leukemia detection.
• Implemented and compared input-based, GAN-based, and loss-based methods.
• Tested approaches on the C-NMC and ALLIDB-2 datasets, known for their imbalanced class distributions.

Main Results:

• Loss-based methods demonstrated superior performance in handling imbalanced classification scenarios for leukemia detection.
• Empirical evidence confirmed the outperformance of loss-based techniques over GAN-based and input-based methods.
• The study provides insights into optimizing deep learning for challenging hematological datasets.

Conclusions:

• Loss-based deep learning methods are highly effective for imbalanced classification tasks in leukemia detection.
• The findings suggest that loss-based approaches offer a promising direction for improving automated diagnostic tools in hematology.
• Further research can build upon these findings to enhance the accuracy and reliability of AI in cancer diagnostics.