Nuclear medicine communications · 2025
Evaluating the efficacy of the ResNet50 deep learning model utilizing thyroid scintigraphy images for predicting the outcomes of initial iodine-131 therapy in patients with Graves' disease.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Lu Lu, Meng Dongyun, Mo Shaozhou, Sun Zeyong, Song Fengyang, Liao Kehua, Huang Wentan
Аннотация
BACKGROUND: Graves' disease (GD), a leading cause of hyperthyroidism, exhibits heterogeneous responses to iodine-131 therapy, underscoring the need for accurate predictive tools. While pertechnetate ( 99m TcO 4- ) thyroid scintigraphy provides essential functional insights, its clinical utility is restricted by limited resolution and the subjective interpretation of results. Deep learning emerges as a promising approach for extracting prognostic features from imaging data.
OBJECTIVE: This research seeks to construct a radiomic model employing deep learning methodologies, utilizing thyroid scintigraphy, to predict the efficacy of initial iodine-131 therapy in patients diagnosed with GD.
METHODS: This retrospective study involved 121 patients diagnosed with GD who underwent pretreatment thyroid scintigraphy. A ResNet50 convolutional neural network was implemented, trained end-to-end on standardized thyroid scintigraphy data to classify treatment outcomes as either cured or not cured. Gradient-weighted class activation mapping (Grad-CAM) was employed to identify region-specific predictive features. Radiomic features extracted from deep learning were further analyzed using least absolute shrinkage and selection operator regression, with a random forest (RF) model serving as the final predictive tool. Performance metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, were evaluated on independent training and validation datasets, using an 8 : 2 split.
RESULTS: The ResNet50 model yielded AUC values of 0.828 [95% confidence interval (CI): 0.740-0.915] and 0.806 (95% CI: 0.656-0.956) for the training and validation datasets, respectively. A set of 16 critical radiomic features, extracted from the global pooling layer of ResNet50, enabled the construction of an RF classifier that exhibited enhanced generalization performance, as evidenced by a test AUC of 0.924. Decision curve analysis substantiated the model's clinical utility across a broad spectrum of threshold probabilities (10-90%). Grad-CAM visualizations revealed that the model predominantly concentrated on regions of thyroid parenchyma associated with treatment-responsive tissue.
CONCLUSION: This study introduces a pioneering deep learning framework that employs thyroid scintigraphy radiomics to predict iodine-131 therapeutic outcomes in GD. By combining ResNet50 for feature extraction with interpretable Grad-CAM localization, this approach advances personalized nuclear medicine strategies and mitigates the 'black box' limitation typically associated with artificial intelligence models. These findings require validation across multicenter cohorts to refine and optimize precision treatment protocols.
OBJECTIVE: This research seeks to construct a radiomic model employing deep learning methodologies, utilizing thyroid scintigraphy, to predict the efficacy of initial iodine-131 therapy in patients diagnosed with GD.
METHODS: This retrospective study involved 121 patients diagnosed with GD who underwent pretreatment thyroid scintigraphy. A ResNet50 convolutional neural network was implemented, trained end-to-end on standardized thyroid scintigraphy data to classify treatment outcomes as either cured or not cured. Gradient-weighted class activation mapping (Grad-CAM) was employed to identify region-specific predictive features. Radiomic features extracted from deep learning were further analyzed using least absolute shrinkage and selection operator regression, with a random forest (RF) model serving as the final predictive tool. Performance metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, were evaluated on independent training and validation datasets, using an 8 : 2 split.
RESULTS: The ResNet50 model yielded AUC values of 0.828 [95% confidence interval (CI): 0.740-0.915] and 0.806 (95% CI: 0.656-0.956) for the training and validation datasets, respectively. A set of 16 critical radiomic features, extracted from the global pooling layer of ResNet50, enabled the construction of an RF classifier that exhibited enhanced generalization performance, as evidenced by a test AUC of 0.924. Decision curve analysis substantiated the model's clinical utility across a broad spectrum of threshold probabilities (10-90%). Grad-CAM visualizations revealed that the model predominantly concentrated on regions of thyroid parenchyma associated with treatment-responsive tissue.
CONCLUSION: This study introduces a pioneering deep learning framework that employs thyroid scintigraphy radiomics to predict iodine-131 therapeutic outcomes in GD. By combining ResNet50 for feature extraction with interpretable Grad-CAM localization, this approach advances personalized nuclear medicine strategies and mitigates the 'black box' limitation typically associated with artificial intelligence models. These findings require validation across multicenter cohorts to refine and optimize precision treatment protocols.