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

Jincao Yao,Yunpeng Wang,Zhikai Lei,Kai Wang,Xiaoxian Li,Jianhua Zhou,Xiang Hao,Jiafei Shen,Zhenping Wang,Rongrong Ru,Yaqing Chen,Yahan Zhou,Chen Chen,Yanming Zhang,Ping Liang,Dong Xu

2024-02-04

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.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The main problem this paper attempts to address is improving the accuracy and interpretability of thyroid nodule diagnosis. Specifically, the paper develops a Computer-Aided Diagnosis (CAD) model based on the Generative Pre-trained Transformer (GPT) architecture—ThyGPT, aimed at assisting radiologists in evaluating the risk of thyroid nodules through semantic-level doctor-patient interactions.

### Main Issues:
1. **Improving Diagnostic Accuracy**: Traditional CAD models, while performing well in some aspects, still suffer from insufficient accuracy. ThyGPT aims to improve diagnostic accuracy by training on large-scale multi-source information (including doctors' diagnostic reports, pathological results, international diagnostic guidelines, research reports, and ultrasound images).
2. **Enhancing Model Interpretability**: Traditional CAD models are often seen as "black boxes," lacking transparency, which leads to a lack of confidence in their diagnostic results among doctors, patients, and healthcare administrators. ThyGPT enhances model interpretability by generating explanatory texts and feature markers, showcasing key factors in the model's decision-making process.
3. **Improving Doctor-Model Interaction**: Existing CAD models mostly provide pattern recognition probabilities and cannot interact further with doctors. ThyGPT allows semantic-level interaction, enabling doctors to observe and consider various intermediate results during the model's analysis process, significantly enhancing doctors' confidence in the CAD model.

### Solutions:
- **Dataset Construction**: The paper uses a dataset of 19,165 thyroid nodule ultrasound cases from Zhejiang Cancer Hospital for model training and validation.
- **Model Design**: ThyGPT is further trained based on the LlaMA2-13B model, including supervised training and language habit training. The model's backbone network uses the Lang-Chain framework, combined with Swin-Transformer and DCNN models for image analysis.
- **Performance Evaluation**: ThyGPT's performance is rigorously evaluated using Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC), sensitivity, and specificity. Experimental results show that radiologists using ThyGPT significantly outperform their peers using traditional methods alone, as well as the model alone.

### Innovations:
1. **First Application of Large-Scale Language Models in Thyroid Nodule Diagnosis**: ThyGPT is the first study to attempt constructing a large-scale language model to assess thyroid nodule risk, integrating multi-source information comprehensively.
2. **Introduction of the AIGC-CAD Concept**: The AIGC-CAD system uses generative large models to generate explanatory texts and feature markers, providing more intuitive display and analysis, better assisting doctors.

In summary, by developing the ThyGPT model, this paper aims to address the issues of accuracy and interpretability in thyroid nodule diagnosis, providing a new direction for future auxiliary diagnostic tools.