Morphological Abnormalities Classification of Red Blood Cells Using Fusion Method on Imbalance Datasets

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Healthcare
• Hematology

Background:

• Red blood cells (RBCs) are vital for oxygen transport, but abnormalities in shape (poikilocytosis) and size (anisocytosis) can indicate serious health issues like anemia and thalassemia.
• Manual microscopic examination of RBCs by hematologists is time-consuming and prone to error.
• Automated classification of RBC morphology is crucial for early and accurate diagnosis of blood-related disorders.

Purpose of the Study:

• To develop and evaluate a deep-learning strategy utilizing Long Short-Term Memory (LSTM) neural networks for the automated classification of abnormal red blood cells.
• To enhance classification accuracy by fusing traditional and high-level features and addressing class imbalance issues.
• To validate the proposed method on diverse datasets and compare its performance against existing benchmark models.

Main Methods:

• A Long Short-Term Memory (LSTM) based neural network was employed for RBC classification.
• Traditional and high-level features were extracted and fused to improve the distinction between abnormal RBC classes.
• A custom loss function was designed by integrating class weights into cross-entropy loss to mitigate the impact of class imbalance.
• The model was trained and evaluated on the Chula-PIC-Lab dataset and a private dataset from Cachar Cancer Hospital and Research Centre (CCHRC).

Main Results:

• The proposed LSTM-based deep-learning approach achieved high average F1-scores and accuracies on both the Chula-PIC-Lab and CCHRC datasets.
• The method demonstrated superior performance compared to benchmark models including Custom CNN, Custom LSTM, Efficient Net-B1, SMOTE, Hybrid NN, and HPKNN.
• The custom loss function effectively addressed class imbalance, leading to more robust classification of underrepresented abnormal RBC types.

Conclusions:

• The developed LSTM-based deep-learning strategy offers an accurate and efficient automated method for classifying abnormal red blood cells.
• This approach has the potential to significantly aid hematologists in the early diagnosis and management of blood-related disorders.
• The feature fusion and custom loss function techniques are effective in improving the performance of deep learning models for medical image classification tasks.