Assessing BRAF status of localized and locally advanced papillary thyroid microcarcinomas using machine learning methods

Papillary thyroid microcarcinomas (PTMCs) account for about 30% of papillary thyroid cancer cases, and the BRAF V600E mutation is one of the driver mutations in the development of papillary thyroid carcinomas. However, traditional molecular diagnostic methods are time-consuming and costly, highlighting the need for alternative approaches, such as machine learning-based analysis of histopathological images. Objective. To evaluate the feasibility of using a Multiple-Instance Learning (MIL) model to predict BRAF status in PTMC patients based on histopathological whole-slide images without explicit region-of-interest annotation. Material and methods. The study included histopathological scans from surgical specimens of 93 PTMC patients (age range 24—72 years, median 48 years) operated between 01.01.2022 and 01.09.2023. BRAF V600E status was determined by Real-Time PCR. For classification, an MIL model (CLAM) with pretrained feature extractors (CTransPath and Virchow) was used. Model performance was assessed using ROC-AUC with 5-fold cross-validation (10 repeats). Results. The BRAF V600E mutation was detected in 62 patients and absent in 31. The highest ROC-AUC (0.83; 95% CI: 0.61—1.00) was achieved using Virchow features with 20% top-attention region filtering. For CTransPath, ROC-AUC was 0.79 (95% CI: 0.60—0.95) with 5% region filtering. Conclusion. The MIL model shows promise for predicting BRAF status in PTMCs from histopathological images, particularly when using Virchow-derived features.