AI Reads Routine Thyroid Slides to Predict BRAF Mutation Before Surgery

Researchers have built an interpretable computational framework that predicts whether a papillary thyroid carcinoma carries the BRAF V600E mutation using only routine Wright-stained fine-needle aspiration cytology slides, the kind of stained preparation many pathology laboratories already produce every day. The study, published in the Journal of Translational Medicine, was designed to address a practical gap: molecular testing for BRAF V600E is not always immediately available at the time clinicians are planning surgery, yet knowing the mutation status preoperatively can support individualized treatment planning in this common thyroid cancer.

The framework, described as a pathomics approach, works on whole-slide images of cytology specimens rather than on tissue sections or genetic assays. Nuclei detected on each slide were segmented and sorted into five cytological channels labeled Tumor, MTumor, Follicle, MFollicle, and Heterocyte. For every channel, the team extracted four families of measurements—deep features learned by neural networks, along with shape, intensity, and texture descriptors—and aggregated them into representations describing each whole slide. This channel-based design means the model does not treat a slide as an undifferentiated pixel field; it explicitly separates nuclear populations that cytologists recognize as biologically distinct.

Model development followed a deliberately exhaustive search. Single-channel models were trained using 5-fold cross-validation repeated 20 times across 117 combinations of feature selectors and classifiers, a procedure intended to test whether any particular channel’s features could predict mutation status on their own and to reduce the risk that reported performance depended on one lucky data split. A fusion model was then constructed by integrating the prediction scores of the five single-channel models, allowing the final decision to draw on complementary information from all nuclear compartments simultaneously.

The retrospective study included 209 patients with histologically confirmed papillary thyroid carcinoma and PCR-confirmed BRAF V600E status, drawn from two centers. Center 1 contributed 156 patients, split into a training set of 109 and an internal test set of 47. Center 2 provided 53 patients for external validation, with one representative slide per patient serving as external test set 1 and additional slides from the same patients serving as external test set 2. This two-center, multi-dataset structure matters because models that look impressive on the data used to build them frequently fail when applied to slides prepared at a different institution.

The fusion model achieved an area under the receiver operating characteristic curve 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. When the independent test datasets were pooled, the AUC was 0.925. An AUC near 0.9 indicates that the model ranks a randomly chosen mutation-positive slide above a randomly chosen mutation-negative slide roughly nine times out of ten, a level of discrimination the authors describe as encouraging across all evaluation sets, including the external data the model never saw during training.

Because a patient’s diagnosis ultimately rests on one slide, the team also asked whether the model’s answer would change if a different slide from the same patient were analyzed. In this same-patient slice-sensitivity analysis, reference slides and additional slides showed moderate-to-good agreement, with a classification agreement of 85.4% and a Cohen’s kappa of 0.654 at the 0.50 decision threshold. Kappa values in this range are generally interpreted as substantial agreement, though the figure also signals that some slide-to-slide variability exists—a realistic caveat for any image-based test that must work on whatever slide reaches the analyzer.

A distinctive feature of the work is its use of Shapley additive analyses to open the model’s reasoning to inspection. These analyses indicated that the fusion model integrated complementary signals across the five cytological channels, with the Follicle channel contributing the largest global share to predictions. Follow-up analysis of that channel identified stable and interpretable features tied to cell identity probability, texture complexity, contour irregularity, and intensity-distribution heterogeneity. In other words, the model’s accuracy appears to rest on measurable properties of nuclear form and staining pattern rather than on opaque correlations that could not be examined or audited.

The authors conclude that the interpretable FNAC-based pathomics framework enabled accurate preoperative prediction of BRAF V600E mutation status in papillary thyroid carcinoma and showed reproducible performance across internal and external validation datasets. They position the approach as a potential decision-support tool for image-based molecular risk assessment—something that could flag likely mutation status from slides already in hand while confirmatory molecular testing proceeds, rather than a replacement for genetic testing itself.

Several boundaries on the finding deserve emphasis. The study was retrospective, and all 209 patients had histologically confirmed cancer with PCR-confirmed mutation status, so the framework was validated on known PTC cases rather than on the full range of thyroid nodules encountered in clinic. Performance was somewhat lower on external test set 2, the additional-slide dataset, and the same-patient agreement, while good, was not perfect. The authors themselves state that the approach warrants prospective multicenter validation before it can be considered for routine use. Within those limits, the study demonstrates that mutation-relevant information is visible in ordinary Wright-stained cytology slides and that machine learning models can extract it in a form that clinicians can, in principle, inspect.

Subject of Research: Development and validation of an interpretable FNAC-based pathomics framework for predicting BRAF V600E mutation status in papillary thyroid carcinoma from Wright-stained cytology whole-slide images.

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

Article References: Guo, J., Yang, D., Zhu, R., Zhou, L., Shen, T., Chen, J., Tan, R., An, P., Zhang, N., Song, L., Tan, T., Hu, L., Lu, C., Yang, F., & Zhao, X. (2026). Development and validation of an interpretable FNAC-based pathomics framework for predicting BRAF V600E mutation status in papillary thyroid carcinoma using Wright-stained slides. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09147-1

Image Credits: AI Generated

DOI: 10.1186/s12967-026-09147-1

Keywords: BRAF V600E, papillary thyroid carcinoma, fine-needle aspiration cytology, pathomics, whole-slide imaging, machine learning, Wright stain, molecular prediction, retrospective study, external validation, Shapley additive analysis, decision support