Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Video

Updated: Jun 27, 2026

A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
04:23

A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images

Published on: April 21, 2023

Optimizing a Multimodal Large Language Model for Ultrasound-Based Thyroid Nodule Malignancy Classification: A

Yu-Hsuan Li1,2,3,4, Yu-Cheng Cheng2,4, Chih-Yun Chang2

  • 1Department of Computer Science & Information Engineering, National Taiwan University, Taipei 106319, Taiwan.

Diagnostics (Basel, Switzerland)
|June 26, 2026
PubMed
Summary

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Pemetrexed potentiates γδ T cell-based immunotherapy in NSCLC through ATM-STING-NF-κB-mediated induction of NKG2D ligands.

International immunopharmacology·2026
Same author

Targeting Galectin-1 with Triptolide Induces Ferroptosis in Oral Squamous Cell Carcinoma.

Cancers·2026
Same author

Circulating tumor cells (CTCs) enumeration and machine-learning based diagnostic biomarkers for breast cancer detection.

BMC cancer·2026
Same author

Climate skepticism, natural hazards, and community vulnerability in COPD mortality among US counties.

BMC pulmonary medicine·2026
Same author

Variability in the urinary albumin-to-creatinine ratio predicts mortality in patients with type 2 diabetes.

Journal of diabetes and its complications·2025
Same author

Disparities in nontuberculous mycobacteria mortality in the United States, 1999-2020.

American journal of respiratory and critical care medicine·2025

A hybrid strategy using GPT-4o demonstrated high specificity for classifying thyroid nodule malignancy, outperforming expert consensus and ATA risk stratification in certain cases. However, its false-negative rate restricts its use as a standalone diagnostic tool.

Area of Science:

  • Artificial Intelligence in Medical Imaging
  • Machine Learning for Diagnostic Classification
  • Multimodal Large Language Models (MLLMs)

Background:

  • Multimodal large language models (MLLMs) show promise for medical image classification tasks.
  • Evaluating optimization strategies is crucial for enhancing MLLM performance in clinical applications.
  • Thyroid nodule malignancy classification presents diagnostic challenges, particularly for nodules with atypia of undetermined significance (AUS).

Purpose of the Study:

  • To evaluate four optimization strategies (text prompting, few-shot learning, fine-tuning, hybrid) for two MLLMs (GPT-4o, Gemini 2.5 Flash-Lite).
  • To assess MLLM performance in ultrasound-based thyroid nodule malignancy classification using public and clinical datasets.
  • To compare the best-performing MLLM strategy against expert endocrinologist consensus and ATA risk stratification in a challenging AUS cohort.
Keywords:
atypia of undetermined significancemultimodal large language modelnodulethyroidultrasound

Related Experiment Videos

Last Updated: Jun 27, 2026

A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
04:23

A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images

Published on: April 21, 2023

Main Methods:

  • Two MLLMs, GPT-4o and Gemini 2.5 Flash-Lite, were evaluated using text prompting, few-shot learning, fine-tuning, and a hybrid approach.
  • Performance was assessed on two public datasets (DDTI, TN5000) and an institutional AUS cohort with surgical pathology.
  • The top strategy in the AUS cohort was compared with endocrinologist consensus and ATA risk stratification.

Main Results:

  • For GPT-4o, the hybrid strategy yielded the highest AUC across all datasets (DDTI: 0.866, TN5000: 0.689, AUS cohort: 0.836).
  • In the AUS cohort, GPT-4o's hybrid strategy achieved higher specificity than expert consensus and ATA stratification (95.1% vs. 70.7%) for high-suspicion nodules, with comparable sensitivity.
  • Gemini 2.5 Flash-Lite did not show performance improvements with advanced strategies over basic text prompting. GPT-4o's hybrid model had a 27.9% false-negative rate in the AUS cohort.

Conclusions:

  • The hybrid strategy significantly improved GPT-4o performance but not Gemini 2.5 Flash-Lite.
  • Optimized GPT-4o demonstrated high specificity in classifying challenging AUS thyroid nodules but its false-negative rate limits standalone diagnostic use.
  • Further validation in larger, prospective, multicenter studies is necessary before clinical implementation of MLLMs for thyroid nodule malignancy classification.