Very Deep Convolutional Neural Networks for Morphologic Classification of Erythrocytes

Area of Science:

• Hematology
• Computational Biology
• Medical Imaging

Background:

• Manual erythrocyte morphology profiling is time-consuming and lacks reproducibility.
• Automated methods using machine learning can improve efficiency and accuracy.
• Convolutional neural networks (CNNs) show promise for image-based classification tasks.

Purpose of the Study:

• To evaluate the performance of deep convolutional neural networks (CNNs) for classifying erythrocyte morphology.
• To assess the accuracy and reliability of a "very deep" CNN model for this task.

Main Methods:

• Erythrocytes were manually classified into 10 morphologic classes using a custom web application.
• A "very deep" CNN architecture with over 150 layers and dense shortcut connections was implemented.
• An ensemble model approach was used for predictions on unseen data.

Main Results:

• The model achieved a harmonic mean of recall and precision of 92.70% and 89.39% on unseen data.
• The overall correct classification frequency was 90.60% across 10 morphologic classes.
• Out of 748 cells in the test set, only 23 misclassifications occurred.

Conclusions:

• "Very deep" CNNs can accurately measure erythrocyte morphology profiles.
• This approach offers a potential clinical tool for faster and more reliable diagnostics.
• Further research can expand classes and optimize performance for clinical integration.