This study aimed to develop and evaluate ThyroSignature, a whole-slide deep-learning model for predicting BRAFV600E status in papillary thyroid carcinoma from routine haematoxylin and eosin histology, and to compare its performance with pathologist assessment of BRAF-like morphology. The model was developed and evaluated using 489 cases with definitive VE1 immunohistochemistry(IHC) results and five-fold stratified cross-validation with out-of-fold predictions. Whole-slide images were tiled, embedded using a pathology foundation model, and aggregated using a transformer and channel-attention multiple-instance-learning architecture. Comparative analyses included 486 cases with pathologist morphology scores: 0, little/no BRAFV600-like atypia; 1, possible BBRAFV600-like atypia; and 2, typical BRAFV600-like atypia.
ThyroSignature achieved an area under the receiver operating characteristic curve of 0.934 (95% confidence interval[CI],0.910–0.956), accuracy of 87.4%, sensitivity of 87.8%, and specificity of 87.0%. Treating morphology score 2 alone as positive yielded a high specificity (92.5%) but low sensitivity (31.8%). Combining scores 1 and 2 as positive increased the sensitivity to 71.3%, but reduced specificity to 69.5%. Agreement between AI and pathologist classification was 59.9% and 71.2% under these respective definitions. When they disagreed, IHC supported the AI model in 172/195 cases (88.2%) under the score-2-only definition and 111/140 cases (79.3%) when scores 1 and 2 were positive. Among 159 cases the pathologist rated as 1, 71.1% were IHC-confirmed BRAF+. TThyroSignature correctly achieved an AUC of 0.886 (95%CI:0.830–0.937) and accuracy of 81.1% (95%CI:74.3–86.5%).
ThyroSignature provided substantially better discrimination and a more balanced sensitivity–specificity profile than morphology alone. It may serve as a reproducible second reader and support triage for confirmatory testing where biomarker access is constrained. Among cases the pathologist rated as possible BRAFV600E-like atypia (score 1), ThyroSignature achieved high accuracy, demonstrating that the model provides clinically actionable predictions precisely where pathologist morphological assessment is unable to render a definitive opinion. External validation is required before clinical implementation.