Risk Assessment of Thyroid Nodules with a Multi-Instance Convolutional Neural Network

Da Yu¹Tingting Song²Yancheng Yu¹Hebin Zhang¹Feng Gao¹Jiacheng Wang³Zirong Wang^4*

• ¹Department of Ultrasound, The Affiliated Hospital of Hangzhou Normal University, Hangzou, China
• ²Department of Radiology, The Fourth People’s Hospital of Harbin, Harbin, China
• ³Department of Computer Science, School of Informatics, Xiamen University, Xiamen, China
• ⁴Department of Radiology, The Affiliated Hospital of Hangzhou Normal University,, Hangzhou, China

Ultrasonography is the primary imaging modality for evaluating thyroid nodules, and artificial intelligence (AI) has advanced the automated diagnosis of thyroid cancer. However, existing AI-assisted methods often suffer from limited diagnostic performance. In this study, we propose a novel multi-instance learning (MIL) convolutional neural network (CNN) model tailored for ultrasound-based thyroid cancer diagnosis. Specifically, the model extracts nodule-level ultrasound features from instance-level images using CNNs, and employs an attention mechanism to assign importance scores and aggregate features across instances. This enables effective feature extraction and localization of key instance features, facilitating risk assessment of thyroid nodules. The dataset consists of ultrasound images from 2000 patients at the Affiliated Hospital of Hangzhou Normal University, collected between 2018 and 2024.The images were divided into training (75%, 1500 patients) and testing (25%, 500 patients) sets. The model's performance was evaluated using metrics, including accuracy, precision, recall, F1-Score, and AUC. To assess the statistical significance of the model's performance relative to other methods, a paired t-test was conducted based on the prediction results. The performance of the model developed in this study was evaluated and compared with popular ultrasound image classification models for 2 thyroid nodules. The model outperformed the other two classification models (accuracy 0.8386 ± 0.0334, 0.7999 ± 0.0188, 0.7839± 0.0267; precision 0.8512 ± 0.0301, 0.9039±0.0154, 0.9267±0.0235; recall0.8427±0.0313, 0.7497±0.0163, 0.6987±0.0249; F1-Score 0.8380±0.0344, 0.8196±0.0178, 0.7967±0.0251; AUC 0.8900±0.0309, 0.8851 ±0.0124, 0.6340±0.0200), , where values are under 95% confidence interval. Statistical analysis showed that the performance differences were statistically significant (p < 0.0001). These results demonstrate the effectiveness and clinical utility of the proposed MIL-CNN framework in non-invasively stratifying thyroid nodule risk, supporting more informed clinical decisions and potentially reducing unnecessary biopsies and surgeries. Codes will be available at https://github.com/rrrr-ops/Thyroid-AI.