AI-Enabled Automated Schistocyte Classification in Peripheral Blood for TMA Auxiliary Diagnosis

Area of Science:

• Artificial Intelligence in Hematology
• Medical Image Analysis
• Digital Pathology

Background:

• Schistocytes are crucial for diagnosing thrombotic microangiopathies (TMA).
• Manual schistocyte identification suffers from inconsistency and interobserver variability.
• AI-driven hematology tools are emerging, but AI for schistocyte classification is underdeveloped.

Purpose of the Study:

• To develop and validate a two-stage AI system for accurate schistocyte segmentation and classification.
• To address the clinical gap in AI-enabled schistocyte analysis for TMA diagnosis.

Main Methods:

• A two-stage AI system was developed, incorporating segmentation and classification models.
• Segmentation models (U-Net, ENet, R2U-Net) were evaluated on 25,067 RBCs.
• Classification models (ResNet-50, Xception) were trained on 28,586 RBCs, including 13,125 schistocytes, using grayscale and RGB images.
• Clinical validation involved 156,784 RBCs from 219 patients, with performance assessed using recall, specificity, precision, and F1 score.

Main Results:

• R2U-Net achieved the highest segmentation performance (recall=0.868, F1=0.881).
• The Xception-RGB model excelled in classification, achieving a weighted F1 score of 0.957 with high precision and recall for schistocyte subtypes.
• Clinical validation demonstrated exceptional reliability, with all weighted metrics reaching 0.998.

Conclusions:

• The developed two-stage AI framework enables precise schistocyte analysis.
• Utilizing RGB images and the Xception model enhances feature representation and diagnostic accuracy.
• This AI approach improves diagnostic reproducibility, aiding in TMA auxiliary diagnosis and clinical decision-making.