Publication
Recommendation of a Surgical Approach for Papillary Thyroid Microcarcinoma Based on Preoperative Ultrasound Features Using a Multiclass Deep Neural Network
Enhui He; Yiran Wang; Xinghao Wang; Zhanxiong Yi; Marcin Grzegorzek; Peipei Yang; Xiangdong hu; Ying Feng; Zhixiang Wang
In: Journal of Medical and Biological Engineering (JMBE), Vol. 45, Pages 378-384, Springer Nature, 6/2025.
Abstract
Purpose
To develop and evaluate a multiclass deep neural network model that utilizes preoperative ultrasound features to recommend the most suitable surgical approach for patients with papillary thyroid microcarcinoma (PTMC), thereby providing empirical guidance for surgical method selection and promoting personalized treatment strategies.
Methods
Data from a cohort of 1,434 cN0 PTMC patients from Beijing Friendship Hospital, Capital Medical University were retrospectively analysed. After dataset standardization, bootstrap resampling and recursive feature elimination (RFE) were performed to identify robust predictors (selected in > 90% iterations). A multiclass deep learning framework integrating ultrasound characteristics, clinical risk factors (e.g., comorbidities), and histopathological variables was developed to classify patients into three surgical categories: total mastectomy (Class 1), subtotal mastectomy (Class 2), and tumour excision/unilateral lobectomy (Class 3). Model performance was evaluated by calculating (ACC), area under the curve (AUC), sensitivity (SEN), and specificity (SPE) values as well as by examining confusion matrices.
Results
The model demonstrated differential performance across various surgical options. Total mastectomy (Class 1) showed moderate accuracy (ACC = 0.73) and a good balance between sensitivity and specificity, with an AUC of 0.83. Subtotal mastectomy (Class 2) exhibited superior performance (ACC = 0.91, AUC = 0.92), albeit with slightly lower sensitivity. The performance of tumour excision or unilateral lobectomy (class 3) was comparable to that of class 1 in terms of the ACC and AUC. The critical variables identified for surgical decision-making included ‘Age’, ‘Comorbidities’, ‘Lesion Size’, ‘Heart Rate’, and ‘Number of Lesions’.
Conclusion
This study demonstrated the feasibility of deep learning models to objectively guide PTMC surgical planning based on preoperative data, thus indicating a critical step towards personalized thyroid cancer management.
