A Hybrid Framework for Red Blood Cell Labeling Using Elliptical Fitting, Autoencoding, and Data Augmentation

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Existing datasets lack population-specific RBC morphology.
• Anemia and thalassemia diagnosis in Southeast Asia require specialized data.

Purpose of the Study:

• To create a local dataset of abnormal RBC morphology for Thailand.
• To support AI-assisted diagnostics in hematology.

Main Methods:

• Collected blood smear samples from six hematological disorders.
• Used convolutional autoencoders and ellipse fitting for feature extraction and quantification.
• Applied expert validation and data augmentation for dataset refinement.

Main Results:

• Developed a dataset of 14,089 high-quality single-cell images.
• Classified RBC morphology into 36 clinically meaningful categories.
• Dataset reflects Southeast Asian population-specific characteristics and diversity.

Conclusions:

• Established a scalable and interpretable dataset integrating computational methods and expert knowledge.
• The dataset is a robust resource for advancing hematology research.
• Facilitates the integration of AI-driven clinical support systems with traditional diagnostics.