WBC-based segmentation and classification on microscopic images: a minor improvement.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Accurate analysis of white blood cells (WBCs) from microscope images is crucial for diagnosing various health conditions.
• Current automated methods for WBC image segmentation and classification lack the required accuracy and robustness.
• The Acute Lymphoblastic Leukemia Image Database (ALL-IDB) is utilized for evaluating the proposed methods.

Purpose of the Study:

• To develop an improved automated method for segmenting and classifying white blood cell (WBC) images.
• To enhance the accuracy and efficiency of identifying healthy and malignant WBCs.
• To address the limitations of existing automated WBC image analysis techniques.

Main Methods:

• A triple thresholding technique was employed for segmenting WBCs from microscope images.
• Morphological opening with a 13x13 kernel was applied to refine segmentation results.
• A Convolutional Neural Network (CNN) model, utilizing transfer learning with InceptionV3, was developed for binary classification of WBCs (healthy vs. malignant).

Main Results:

• The WBC segmentation method achieved 90.45% accuracy, 83.81% SSIM, and 76.25% Dice coefficient.
• The fine-tuned CNN classifier demonstrated high accuracy (96.15% on test set) and precision (>96%) for classifying WBCs.
• The proposed triple thresholding method outperformed K-means clustering for segmentation on smaller datasets.

Conclusions:

• The developed automated method significantly improves the accuracy of WBC image segmentation and classification.
• Transfer learning with a pre-trained InceptionV3 model enhances the classifier's flexibility and performance.
• This approach offers a computationally efficient and accurate tool for potential clinical applications in hematology.