Hybrid adversarial-discriminative network for leukocyte classification in leukemia

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Machine Learning in Healthcare

Background:

• Leukemia diagnosis relies heavily on accurate white blood cell classification.
• Current methods may face challenges in extracting discriminative features for precise diagnosis.

Purpose of the Study:

• To develop and evaluate a hybrid adversarial residual network with Support Vector Machine (SVM) for enhanced white blood cell classification.
• To improve the accuracy of human peripheral white blood cell classification by integrating cell segmentation and fine-grained features.

Main Methods:

• Utilized an adversarial residual network for segmenting cells and nuclei.
• Extracted Convolutional Neural Network (CNN) and Histogram of Oriented Gradient (HOG) features from segmented cell nuclei.
• Combined CNN and HOG features and employed a linear Support Vector Machine (SVM) for classifying six types of white blood cells.

Main Results:

• The hybrid approach combining CNN and HOG features achieved 95.93% accuracy, 94.57% specificity, and 96.11% sensitivity.
• Individual CNN features yielded 94.41% accuracy, while HOG features resulted in 85.00% accuracy.
• Evaluated on 5000 leukocyte images, demonstrating superior performance of the combined feature approach.

Conclusions:

• A novel hybrid adversarial-discriminative network effectively classifies microscopic leukocyte images.
• The proposed method enhances classification accuracy, potentially reducing diagnostic time and workload for clinicians.
• This approach offers a valuable tool for efficient and accurate clinical diagnosis of leukemia.