Work

Ontology-guided neuro-symbolic convNeXtV2–swin transformer for explainable cytological classification of papillary thyroid carcinoma

Year: 2026

Type: article

Abstract: The cytological assessment of papillary thyroid carcinoma (PTC) relies on microscopic evaluation of cellular and nuclear morphology. Subtle cytomorphological differences, interobserver variability, and specimen heterogeneity can complicate interpretation. This study developed an explainable hybrid ConvNeXtV2–Swin Transformer framework that integrates fine-grained cytological representation with broader spatial-context modeling for PTC classification. ConvNeXtV2 provides fine-grained nuclear and cellular representations, whereas Swin Transformer captures broader contextual relationships across cytological regions. These complementary representations support classification and subsequent pathology-informed interpretation of nuclear, architectural, and stromal evidence. Interpretability combines Biomarker-Guided Attention Modeling, Gradient-weighted Class Activation Mapping (Grad-CAM)-style quantitative attribution analysis, and ontology-guided post-classification semantic interpretation. On the independent test set, the hybrid framework achieved 96.19% accuracy, 96.16% sensitivity, and 96.16% specificity. Precision was 95.27%, and the area under the receiver operating characteristic curve (ROC-AUC) reached 0.9702. External testing achieved 91.4% accuracy, 91.0% sensitivity, 92.0% specificity, and 94.0% precision. Quantitative attribution analysis showed greater activation heterogeneity and stronger localized responses in PTC tissue. Non-PTC tissue showed broader and more uniform responses associated with preserved follicular morphology. The neuro-symbolic component subsequently organized compartment-level responses according to predefined thyroid cytopathology relationships. These outputs provide pathology-informed computational interpretation without constituting independent biomarker detection or expert-equivalent pathological reasoning.

Source: Scientific Reports

DOI: 10.1038/s41598-026-74498-0