RET-US Study - Ultrasound-Based Prediction of RET Alterations and Lateral-Neck Metastasis in Thyroid Cancer

Overview

Why is this study being done? RET gene alterations occur in only 5-10 % of papillary thyroid cancers, but they can change how surgeons treat the disease. Gene testing is costly and not always performed, so many RET-positive tumours are missed. Researchers have built a computer program (artificial-intelligence or "AI" model) that reads routine thyroid ultrasound images and predicts whether the tumour carries a RET alteration and whether the cancer has already spread to lymph-nodes in the side of the neck. What will happen in this study? About 800 adults who are scheduled for thyroid-cancer surgery will take part. Each participant will: * have a standard pre-operative ultrasound exam (no extra scanning time), * give a routine fine-needle sample for a 14-gene panel test (results in 24 h), and * allow the AI model to analyse the ultrasound images in the background. Doctors making treatment decisions will not see the AI result. After surgery, the research team will compare the AI predictions with the gene-panel result and the final pathology report. Main goal: To find out how accurately the AI model detects RET alterations. Secondary goals: To measure the model's ability to predict lymph-node spread, and to compare costs between ultrasound-only prediction and full gene testing. Benefits and risks: Participants will receive the current standard of care; there is no added risk beyond the usual ultrasound and needle biopsy. The study could lead to faster, less expensive ways to identify high-risk thyroid cancers in the future.

Background RET rearrangements or point mutations drive a minority of papillary thyroid carcinomas (PTC) yet are associated with aggressive behaviour and may qualify patients for selective RET inhibitors. Because of low prevalence, RET testing is often omitted, resulting in under-recognition. Recent work shows that high-resolution ultrasound contains radiomic signatures linked to tumour genotypes. A deep-learning model (EfficientNet-B3 backbone with dual segmentation + multi-label heads) was trained on 1 000 retrospectively collected cases, including 74 RET-positive tumours augmented with GAN-based synthetic images, achieving an AUC of 0.87 for RET prediction in internal cross-validation. Objectives Primary: validate the AI model's area under the receiver-operating characteristic curve (AUC) for RET alteration detection in a prospective cohort. Secondary: (i) sensitivity/specificity for RET; (ii) accuracy for predicting lateral-neck (pN1b) metastasis; (iii) incremental cost per correct RET diagnosis; (iv) concordance between AI probability score and lymph-node burden. Design Single-arm, prospective observational cohort (n = 800). Consecutive eligible patients will undergo: (1) routine pre-operative thyroid ultrasound; (2) upload of DICOM files to a cloud inference server; (3) rapid 14-gene next-generation sequencing panel on FNA or paraffin tissue (includes RET fusions KIF5B, CCDC6, NCOA4 and point mutations M918T, V804). Surgeons remain blinded to AI output. Surgical specimens provide ground truth for pN staging. Data captured in REDCap; statistical analysis uses DeLong test for AUC and McNemar test for paired accuracy. Eligibility Adults 18-75 y with radiologically suspected PTC, planned thyroidectomy, and consent for gene testing. Exclusions: re-operative neck, medullary/anaplastic carcinoma, pregnancy, eGFR \< 30 mL min-¹ 1.73 m-². Sample Size With expected RET prevalence 6 % and target AUC ≥ 0.80 vs null 0.50, 800 cases provide 90 % power (α = 0.05). Ethics \& Oversight IRB approved; minimal-risk diagnostic study. Ultrasound and FNA are standard-of-care; AI inference uses de-identified images. Results will be disseminated via peer-reviewed journals and conference presentations.