Decoupling Detection and Classification to Improve Morphological Phenotype Analysis of Sickle Red Blood Cells in...

Area of Science:

• Biomedical Image Analysis
• Artificial Intelligence in Hematology
• Computational Pathology

Background:

• Red blood cell (RBC) morphology analysis is crucial for sickle cell disease (SCD) phenotyping.
• Existing AI models struggle with full-scope images due to cell density and diverse morphologies.
• Accurate detection and fine-grained classification are needed for comprehensive RBC analysis.

Purpose of the Study:

• To develop an end-to-end computational framework for RBC detection and classification in microscopy images.
• To address limitations of single-step AI models in handling complex, real-world microscopic data.
• To classify RBCs into five key morphological categories: discocytes, echinocytes, elongated/sickle-shaped cells, granular cells, and reticulocytes.

Main Methods:

• Evaluation of advanced detection models (YOLO, DETR) for cell localization.
• Implementation of a two-step framework: YOLO-based detection/cropping followed by DenseNet121 classification.
• Fine-tuning a DenseNet121 ensemble classifier for accurate morphological categorization.

Main Results:

• The proposed framework achieved a detection F1-score of 0.9661 and classification F1-score of 0.9708.
• Overall classification accuracy reached 97.06%.
• Significant improvements in macro-average F1-score (+0.1675) and minority class performance compared to single-step models.

Conclusions:

• A hybrid, two-step AI framework effectively combines general detection with specialized classification for RBC morphology analysis.
• This approach enhances accuracy and efficiency for scientific and clinical image analysis in SCD.
• The framework offers a practical strategy for adapting AI to complex biomedical imaging tasks.