Development and validation of an interpretable FNAC-based pathomics framework for predicting BRAF V600E mutation status in papillary thyroid carcinoma using Wright-stained slides

Abstract Background Preoperative prediction of BRAF V600E mutation status in papillary thyroid carcinoma (PTC) may support individualized treatment planning, but molecular testing is not always immediately available. This study developed and validated an interpretable pathomics framework for predicting BRAF V600E status from routine Wright-stained fine-needle aspiration cytology (FNAC) whole-slide images. Methods This retrospective study included 209 patients with histologically confirmed PTC and PCR-confirmed BRAF V600E status from two centers. Center 1 contributed 156 patients, which were divided into a training set ( n = 109) and an internal test set ( n = 47). Center 2 provided 53 patients for external validation, including one representative slide per patient as external test set 1 and additional slides as external test set 2. Nuclei were segmented and classified into five cytological channels: Tumor, MTumor, Follicle, MFollicle, and Heterocyte. For each channel, deep, shape, intensity, and texture features were extracted and aggregated into whole slide image (WSI)-level representations. Single-channel models were developed using 5-fold cross-validation with 20 repeats across 117 feature selector-classifier combinations, and a fusion model was constructed by integrating the prediction scores of the five single-channel models. Results The fusion model achieved encouraging discrimination across datasets, with area under the receiver operating characteristic curve (AUC) of 0.947 in the cross-validation training set, 0.933 in the internal test set, 0.962 in external test set 1, and 0.889 in external test set 2. In the pooled independent test datasets, the fusion model achieved an AUC of 0.925. Same-patient slice-sensitivity analysis showed moderate-to-good agreement between reference and additional slides, with a classification agreement of 85.4% and Cohen’s kappa of 0.654 at the 0.50 threshold. Shapley additive analyses indicated that the fusion model integrated complementary signals across cytological channels, with the Follicle channel showing the largest global contribution. Further Follicle-channel analysis identified stable and interpretable features related to cell identity probability, texture complexity, contour irregularity, and intensity-distribution heterogeneity. Conclusions An interpretable FNAC-based pathomics framework enabled accurate preoperative prediction of BRAF V600E mutation status in PTC and showed reproducible performance across internal and external validation datasets. This approach may provide a practical decision-support tool for image-based molecular risk assessment, warranting prospective multicenter validation.