Recovering BRAF Signal from Individual Tumor Patches in Papillary Thyroid Carcinoma H&E Whole-Slide Images

YongHun Lee, Jiwon Song, Juhyeong KI, YongKyung Oh, HaYoung Lee, EunKyung Lee, Jimin Lee

Keywords: Whole-Slide Imaging, BRAF Mutation, Representation Learning, Computational Pathology

Registration Requirement: Yes

Abstract: BRAF V600E mutation has important diagnostic and prognostic implications in papillary thyroid carcinoma (PTC), and predicting its status from H&E whole-slide images could replace ancillary testing for BRAF assessment in selected settings. In this study, we found that tumor patches from BRAF-wildtype and BRAF-mutant cases were distinguishable at the patch level, and that this difference remained evident after aggregation to the slide level. To examine this, we performed a controlled comparison of patch-level feature representations within a tumor-focused pipeline using a curated PTC H&E WSI cohort of 665 patients with matched BRAF labels. Pathology foundation models (UNI2 and Virchow2) showed higher performance than an ImageNet-pretrained ResNet50 under identical downstream settings. Overall, the results support the presence of local morphologic patterns linked to BRAF status in PTC H&E morphology.

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 77