An explainable AI-based blood cell classification using optimized convolutional neural network.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Accurate classification of white blood cells (WBCs) is critical for disease diagnosis.
• Existing methods may face challenges with noise and feature extraction in blood cell images.

Purpose of the Study:

• To develop an enhanced convolutional neural network (CNN) for precise blood cell detection and classification.
• To improve the interpretability and transparency of the CNN model for medical applications.

Main Methods:

• Utilized image pre-processing techniques (padding, thresholding, erosion, dilation, masking) to enhance blood cell images.
• Optimized CNN architecture and hyperparameters for improved performance.
• Compared the proposed model against transfer learning models (Inception V3, MobileNetV2, DenseNet201).
• Employed SHAP, LIME, Grad-CAM, and Grad-CAM++ for model interpretability and visualization.

Main Results:

• The proposed CNN model achieved a testing accuracy of 99.12%, precision of 99%, and F1-score of 99%.
• Outperformed Inception V3, MobileNetV2, and DenseNet201 in blood cell classification.
• Grad-CAM++ demonstrated slightly superior performance over Grad-CAM in localization.
• Interpretability techniques provided insights into model decision-making.

Conclusions:

• The enhanced CNN offers a highly accurate and interpretable solution for blood cell classification.
• The model's integration into an end-to-end system (web and Android) facilitates practical use by medical professionals.
• This advancement supports more reliable disease diagnosis through automated cell analysis.