Value of salivary gland imaging based on deep learning and Delta radiomics in evaluation of salivary gland injury following 131I therapy post thyroid cancer surgery
10.3760/cma.j.cn321828-20230919-00052
- VernacularTitle:基于深度学习及Delta影像组学的唾液腺显像在甲状腺癌术后及 131I治疗后唾液腺损伤评估中的价值
- Author:
Yulong ZENG
1
;
Zhao GE
;
Weixia CHONG
;
Jie QIN
;
Biyun MO
;
Wei FU
Author Information
1. 桂林医学院附属医院核医学科，桂林 541000
- Keywords:
Thyroid neoplasms;
Radiotherapy;
Iodine radioisotopes;
Radiation injuries;
Salivary glands;
Radiomics;
Deep learning
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2024;44(2):68-73
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the value of salivary gland imaging based on deep learning and Delta radiomics in assessing salivary gland injury after 131I treatment in post-thyroidectomy thyroid cancer patients. Methods:A retrospective analysis on 223 patients (46 males, 177 females, age(47.7±14.0) years ) with papillary thyroid cancer, who underwent total thyroidectomy and 131I treatment in Affiliated Hospital of Guilin Medical University between December 2019 and January 2022, was conducted. All patients underwent salivary gland 99Tc mO 4- imaging before and after 131I therapy. The patients were categorized according to salivary gland function based on 99Tc mO 4- imaging results (normal salivary gland vs salivary gland injury), and divided into training and test sets in a ratio of 7∶3. A ResNet-34 neural network model was trained using images at the time of maximum salivary gland radioactivity and those based on background radioactivity counts for structured image feature data. The Delta radiomics approach was then used to subtract the image feature values of the two periods, followed by feature selection through t-test, correlation analysis, and the least absolute shrinkage and selection operator( LASSO) algorithm, to develop logistic regression (LR), support vector machine (SVM), and K-nearest neighbor (KNN) predictive models. The diagnostic performance of 3 models for salivary gland function on the test set was compared with that of the manual interpretation. The AUCs of the 3 models on the test set were compared (Delong test). Results:Among the 67 cases of the test set, the diagnostic accuracy of 3 physicians were 89.6%(60/67), 83.6%(56/67), and 82.1%(55/67) respectively, with the time required for diagnosis of 56, 74 and 55 min, respectively. The accuracies of LR, SVM, and KNN models were 91.0%(61/67), 86.6%(58/67), and 82.1%(55/67), with the required times of 12.5, 15.3 and 17.9 s, respectively. All 3 radiomics models demonstrated good classification and predictive capabilities, with AUC values for the training set of 0.972, 0.965, and 0.943, and for the test set of 0.954, 0.913, and 0.791, respectively. There were no significant differences among the AUC values for the test set ( z values: 0.72, 1.18, 1.82, all P>0.05). Conclusion:The models based on deep learning and Delta radiomics possess high predictive value in assessing salivary gland injury following 131I treatment after surgery in patients with thyroid cancer.
