SegMWB: A lightweight deep learning framework for microscopic image classification.

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate white blood cell (WBC) classification is crucial for diagnosing hematological malignancies.
• Automated systems are needed for early and efficient disease detection.

Purpose of the Study:

• To propose a novel deep learning framework, SegMWB, for automated WBC classification from microscopic images.
• To enhance the efficiency and accuracy of WBC classification for early diagnosis of blood disorders.

Main Methods:

• Developed a deep learning framework (SegMWB) with pre-processing, nucleus segmentation, and classification phases.
• Utilized a customized nucleus segmentation algorithm and the SegMWB-Net architecture with block-wise convolutional layers.
• Employed batch normalization to improve convergence and reduce overfitting.

Main Results:

• Achieved high classification accuracies: 96.54% (Raabin-WBC), 98.37% (Peripheral Blood Cell), and 98.70% (LISC).
• Demonstrated improved WBC classification with lower computational complexity compared to transfer-learning models.
• The segmentation-assisted lightweight framework showed competitive performance across diverse datasets.

Conclusions:

• The SegMWB framework offers an efficient and accurate approach for automated WBC classification.
• This deep learning model shows potential for early diagnosis of hematological conditions.
• Further validation on clinical datasets is recommended for practical deployment.