A Multimodal Large Language Model as an End-to-End Classifier of Thyroid Nodule Malignancy Risk: Usability Study
Bibliographic record
Abstract
Background: Thyroid nodules are common, with ultrasound imaging as the primary modality for their assessment. Risk stratification systems like the American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) have been developed but suffer from interobserver variability and low specificity. Artificial intelligence, particularly large language models (LLMs) with multimodal capabilities, presents opportunities for efficient end-to-end diagnostic processes. However, their clinical utility remains uncertain. Objective: This study evaluates the accuracy and consistency of multimodal LLMs for thyroid nodule risk stratification using the ACR TI-RADS system, examining the effects of model fine-tuning, image annotation, prompt engineering, and comparing open-source versus commercial models. Methods: In total, 3 multimodal vision-language models were evaluated: Microsoft's open-source Large Language and Visual Assistant (LLaVA) model, its medically fine-tuned variant (Large Language and Vision Assistant for bioMedicine [LLaVA-Med]), and OpenAI's commercial o3 model. A total of 192 thyroid nodules from publicly available ultrasound image datasets were assessed. Each model was evaluated using 2 prompts (basic and modified) and 2 image scenarios (unlabeled vs radiologist-annotated), yielding 6912 responses. Model outputs were compared with expert ratings for accuracy and consistency. Statistical comparisons included Chi-square tests, Mann-Whitney U tests, and Fleiss' kappa for interrater reliability. Results: Overall, 88.4% (6110/6912) of responses were valid, with the o3 model producing the highest validity rate (2273/2304, 98.6%), followed by LLaVA (2108/2304, 91.5%) and LLaVA-Med (1729/2304, 75%; P<.001). The o3 model demonstrated the highest accuracy overall, achieving up to 57.3% accuracy in Thyroid Imaging Reporting and Data System (TI-RADS) classification, although still remaining suboptimal. Labeled images improved accuracy marginally in nodule margin assessment only when evaluating LLaVA models (407/768, 53% to 447/768, 58.2%; P=.04). Prompt engineering improved accuracy for composition (649/1,152, 56.3% vs 483/1152, 41.9%; P<.001), but significantly reduced accuracy for shape, margins, and overall classification. Consistency was the highest with the o3 model (up to 85.4%), but was comparable for LLaVA and significantly improved with image labeling and modified prompts across multiple TI-RADS categories (P<.001). Subgroup analysis for o3 alone showed prompt engineering did not affect accuracy significantly but markedly improved consistency across all TI-RADS categories (up to 97.1% for shape, P<.001). Interrater reliability was consistently poor across all combinations (Fleiss' kappa<0.60). Conclusions: The study demonstrates the comparative advantages and limitations of multimodal LLMs for thyroid nodule risk stratification. While the commercial model (o3) consistently outperformed open-source models in accuracy and consistency, even the best-performing model outputs remained suboptimal for direct clinical deployment. Prompt engineering significantly enhanced output consistency, particularly in the commercial model. These findings underline the importance of strategic model optimization techniques and highlight areas requiring further development before multimodal LLMs can be reliably used in clinical thyroid imaging workflows.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.019 | 0.066 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".