https://atlas.rsna.org/schemas/2025-11/model.json

Cine-CNNTrans: Deep learning–based risk stratification system for thyroid nodules using ultrasound cine-clip images

https://dx.doi.org/10.1148/ryai.210174

1.0

Supported in part by Yahoo Faculty Research and Engagement Program; NIH awards (PADLY, PADPM, PAEPI, PAFFL, PAWAO, PCOXK, PCQUD, PCRMR); NIH NCI F99/K00 (K00CA234954); National Science Foundation grant(s); additional support acknowledged from ACR, AstraZeneca, Philips (per author disclosures).

Institutional review board–approved, HIPAA-compliant retrospective study; informed consent waived.

Hybrid deep learning model combining a CNN feature extractor (MobileNetV2 pretrained on ImageNet) operating on adjacent-frame stacks (AF-stacks) with a transformer encoder for sequence modeling of framewise embeddings; weakly supervised training at nodule level; focal loss with weighted oversampling to address class imbalance.

Source code and model weights: https://github.com/tarakapoor/thyroid_deep_learning. Dataset (US cine clips, segmentations, and labels): https://stanfordaimi.azurewebsites.net/datasets/a72f2b02-7b53-4c5d-963c-d7253220bfd5

Improves diagnostic performance for thyroid nodule malignancy risk estimation using US cine images and increases specificity of ACR TI-RADS management recommendations, reducing unnecessary biopsies while maintaining sensitivity.

Clinical decision support systems; secondary risk stratification to revise management recommendations.

Operating point chosen per cross-validation fold using a weighted Youden index maximizing (2/3 × sensitivity + 1/3 × specificity − 1), placing 2× weight on sensitivity; outputs binarized into low vs high risk.

Computer-aided decision support; provides malignancy risk score and suggested one-level adjustment to ACR TI-RADS recommendation; requires prior ROI segmentation.

Risk stratification of thyroid nodules in adult patients undergoing thyroid ultrasound with cine sweeps in a clinical radiology setting to inform follow-up vs biopsy recommendations in conjunction with ACR TI-RADS.

Ultrasound cine-clip images of thyroid nodules (grayscale, transverse/longitudinal sweeps) with radiologist-drawn ROI segmentations; AF-stacks of three adjacent frames cropped to lesion bounding box with 5-pixel buffer.

Acquire standard grayscale thyroid US cine clips (12–18 MHz) with transverse/longitudinal sweeps; create ROI segmentations for nodules; generate AF-stacks; run Cine-CNNTrans to obtain continuous malignancy risk and binarize using predefined operating point; optionally revise ACR TI-RADS by one level (upgrade from no-biopsy to follow-up if high risk; downgrade from biopsy to follow-up if low risk; no change if original follow-up).

Single-center, retrospective dataset of 192 nodules (17 malignant); strong class imbalance (175 benign:17 malignant); cross-validation only (no held-out or external validation); manual segmentation required; ground truth from fine-needle aspiration biopsies may be subject to sampling error; operating point determined per fold; potential domain shift across scanners/sites not evaluated.

Use as a secondary tool to revise ACR TI-RADS management by one level: upgrade no-biopsy to follow-up when high risk; downgrade biopsy to follow-up when low risk; do not change follow-up recommendations.

Fivefold cross-validation with fixed folds (patient-level separation); code and pretrained weights publicly available to reproduce training/evaluation.