Published in:
01-10-2019 | Computed Tomography | Imaging Informatics and Artificial Intelligence
Application of deep learning to the diagnosis of cervical lymph node metastasis from thyroid cancer with CT
Published in: European Radiology | Issue 10/2019
Login to get accessAbstract
Purpose
To develop a deep learning–based computer-aided diagnosis (CAD) system for use in the CT diagnosis of cervical lymph node metastasis (LNM) in patients with thyroid cancer.
Methods
A total of 995 axial CT images that included benign (n = 647) and malignant (n = 348) lymph nodes were collected from 202 patients with thyroid cancer who underwent CT for surgical planning between July 2017 and January 2018. The datasets were randomly split into training (79.0%), validation (10.5%), and test (10.5%) datasets. Eight deep convolutional neural network (CNN) models were used to classify the images into metastatic or benign lymph nodes. Pretrained networks were used on the ImageNet and the best-performing algorithm was selected. Class-specific discriminative regions were visualized with attention heatmap using a global average pooling method.
Results
The area under the ROC curve (AUROC) for the tested algorithms ranged from 0.909 to 0.953. The sensitivity, specificity, and accuracy of the best-performing algorithm were all 90.4%, respectively. Attention heatmap highlighted important subregions for further clinical review.
Conclusion
A deep learning–based CAD system could accurately classify cervical LNM in patients with thyroid cancer on preoperative CT with an AUROC of 0.953. Whether this approach has clinical utility will require evaluation in a clinical setting.
Key Points
• A deep learning–based CAD system could accurately classify cervical lymph node metastasis. The AUROC for the eight tested algorithms ranged from 0.909 to 0.953.
• Of the eight models, the ResNet50 algorithm was the best-performing model for the validation dataset with 0.953 AUROC. The sensitivity, specificity, and accuracy of the ResNet50 model were all 90.4%, respectively, in the test dataset.
• Based on its high accuracy of 90.4%, we consider that this model may be useful in a clinical setting to detect LNM on preoperative CT in patients with thyroid cancer.