An Early Thyroid Screening Model Based on Transformer and Secondary Transfer Learning for Chest and Thyroid CT Images

Introduction: Thyroid cancer is a common malignant tumor, and early diagnosis and timely treatment are crucial to improve patient prognosis. With the increasing use of enhanced CT scans, a new opportunity for early thyroid cancer screening has emerged. However, existing CT-based models face challenges due to limited datasets, small sample sizes, and high noise. Methods: To address these challenges, we collected enhanced CT scan image data from 240 patients in Guangdong and Xinjiang, China, and established a CT dataset for early thyroid cancer screening. We propose a deep learning model, the DVT model, which combines transformer DNN and transfer learning techniques to integrate time series data and address small sample sizes and high noise. Results: The experimental results show that the DVT model achieves a prediction accuracy of 0.96, AUROC of 0.97, specificity of 1, and sensitivity of 0.94. These results indicate that the DVT model is a highly effective tool for early thyroid cancer screening. Conclusion: The DVT model has the potential to assist clinicians in identifying potential thyroid cancer patients and reducing patient expenses. Our study provides a new approach to thyroid cancer screening using enhanced CT scans and demonstrates the effectiveness of deep learning techniques in addressing the challenges associated with CT-based models.