[Paper Review] AI-Generated Content Enhanced Computer-Aided Diagnosis Model for Thyroid Nodules: A ChatGPT-Style Assistant

Jincao Yao, Yunpeng Wang+14 more|arXiv (Cornell University)|Feb 4, 2024

Artificial Intelligence in Healthcare and Education12 citations

TL;DR

ThyGPT, an AI-generated content–enhanced CAD model inspired by ChatGPT, aids radiologists in thyroid nodule risk assessment and improves diagnostic performance, especially for junior clinicians.

ABSTRACT

An artificial intelligence-generated content-enhanced computer-aided diagnosis (AIGC-CAD) model, designated as ThyGPT, has been developed. This model, inspired by the architecture of ChatGPT, could assist radiologists in assessing the risk of thyroid nodules through semantic-level human-machine interaction. A dataset comprising 19,165 thyroid nodule ultrasound cases from Zhejiang Cancer Hospital was assembled to facilitate the training and validation of the model. After training, ThyGPT could automatically evaluate thyroid nodule and engage in effective communication with physicians through human-computer interaction. The performance of ThyGPT was rigorously quantified using established metrics such as the receiver operating characteristic (ROC) curve, area under the curve (AUC), sensitivity, and specificity. The empirical findings revealed that radiologists, when supplemented with ThyGPT, markedly surpassed the diagnostic acumen of their peers utilizing traditional methods as well as the performance of the model in isolation. These findings suggest that AIGC-CAD systems, exemplified by ThyGPT, hold the promise to fundamentally transform the diagnostic workflows of radiologists in forthcoming years.

Motivation & Objective

• Bridge interaction and understanding gaps between clinicians and CAD models by adding explainable AI-generated content.
• Build a large language model (ThyGPT) trained on multi-source thyroid data to assess nodule risk.
• Enable semantic-level communication and feature contribution marking to support clinical decision-making.

Proposed method

• Base architecture on LLaMA2-13B and fine-tune with thyroid-specific instructions and reports.
• Integrate an image analysis module using Swin-Transformer and DCNN for feature extraction and heatmap generation.
• Fuse image analysis with nodule recognition in a Lang-Chain–based framework for interactive explanations.
• Train on a large Zhejiang Cancer Hospital dataset (19,165 patients, 12,073 reports) and evaluate with ROC, AUC, sensitivity, and specificity.
• Use data augmentation (rotation ±10°, random crop, 80–120% scaling) to improve robustness across ultrasound machines.

Experimental results

Research questions

• RQ1 Can a ChatGPT-style LLM provide transparent rationale and feature contributions for thyroid nodule risk assessment?
• RQ2 Does integrating AIGC-CAD with radiologists improve diagnostic accuracy compared with radiologists alone or standalone CAD?
• RQ3 How does ThyGPT affect diagnostic performance across junior and senior radiologists on independent test sets?
• RQ4 What is the role of explainable outputs (heatmaps and textual rationale) in clinician trust and decision-making?

Key findings

• ThyGPT achieved an AUC of 0.909 on Test Set 1.
• Junior physicians’ sensitivity increased by 27.1–28.5% and specificity by 19.4–21.3% when aided by ThyGPT.
• Senior physicians’ sensitivity increased by 11.1–14.3% and specificity by 9.2–10.6% when aided by ThyGPT.
• ThyGPT alone showed sensitivities of 84.7–86.2% and specificities of 87.0–88.3% on independent test sets.
• Radiologists plus ThyGPT sometimes surpassed the standalone AI model and rivaled or exceeded senior physicians in some metrics.
• ThyGPT helped reduce inter-observer variability and provided transparent, rationale-based diagnostic support.