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

Abstract

Background: Schistocytes are critical morphological markers for thrombotic microangiopathy (TMA) diagnosis. Manual identification is hampered by inconsistent standards and high interobserver variability, compromising accuracy. Although AI has advanced in hematology, AI-enabled schistocyte classification remains under-studied, creating a clinical gap.

Methods: We constructed a two-stage AI system for segmentation and classification. Three segmentation models (U-Net, ENet, and R2U-Net) were assessed on 25 067 RBCs from 183 patients. ResNet-50 and Xception were trained on 28 586 RBCs, including 13 125 ICSH-classified schistocytes, using grayscale and RGB images. Clinical validation was performed on 156 784 RBCs from 219 patients. Performance was evaluated using recall, specificity, precision, and F₁ score.

Results: R2U-Net achieved the best segmentation (recall = 0.868, F₁ = 0.881, mIoU = 0.807). The Xception-RGB model performed best in classification (weighted F₁ = 0.957), with high precision and recall for schistocyte subtypes. Clinical validation showed excellent reliability with all weighted metrics at 0.998.

Conclusions: This two-stage framework enables accurate schistocyte analysis with better feature representation using RGB images and Xception. It improves diagnostic accuracy and reproducibility, supporting TMA auxiliary diagnosis and clinical decision-making.