Classification of white blood cells using weighted optimized deformable convolutional neural networks.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Machine learning (ML) is utilized for white blood cell (WBC) classification.
• Current ML algorithms face challenges due to a lack of gold-standard datasets and implementation issues.
• Improving the accuracy and robustness of WBC classification is crucial for clinical diagnostics.

Purpose of the Study:

• To introduce a novel deep learning model for enhanced WBC classification.
• To address the limitations of existing ML methods in handling low-resolution and noisy data.
• To evaluate the proposed model against established ML algorithms.

Main Methods:

• A two-module weighted optimized deformable convolutional neural network (TWO-DCNN) was developed.
• The model incorporates two-module transfer learning and deformable convolutional (DC) layers for improved robustness.
• Performance was validated against VGG16, VGG19, Inception-V3, ResNet-50, SVM, MLP, DT, and RF on custom and public datasets.

Main Results:

• TWO-DCNN achieved superior performance across metrics including precision, recall, F1-score, and AUC.
• Specifically, precisions, recalls, and F1-scores reached 95.7%, 94.5%, and 91.6% for low-resolution/noisy and BCCD datasets, respectively.
• Area under the curve (AUC) values of 0.98, 0.97, and 0.95 were recorded for the respective datasets.

Conclusions:

• The proposed TWO-DCNN demonstrates excellent performance in WBC classification, particularly for challenging low-resolution and noisy datasets.
• Accurate feature extraction and optimized network weights contribute to the model's effectiveness.
• TWO-DCNN presents a viable alternative for clinical WBC classification applications.