(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 8
Purpose: To develop deep learning (DL) models for prediction of the malignancy grade of parotid gland cancer (PGC) using three DL architectures with transfer learning using preoperative MR images.
Methods: We retrospectively collected the preoperative MR images of 42 PGC patients with 21 high- and 21 intermediate-/low-malignancy grades. A total of 849 PGC images were cropped from the slice images of preoperative MR images. High- versus intermediate- and low-malignancy grades were predicted using three convolutional neural networks (CNNs) of AlexNet, GoogLeNet, and Inception_v3, which have been pre-trained for classifying million natural color (non-medical) images into 1,000 categories. The PGC images were rescaled by using an optimal input image size before inputting to three models. Last three layers of the pre-trained CNNs were replaced with new three layers of a fully connected layer, a softmax layer, and an output layer for the prediction task, and they are fine-tuned with 1,003 iterations and a learning rate of 0.0001. The images were divided into 70% (594 images) for the fine-tuning and 30% (355 images) for the test of the models. The test was performed 50 times with taking into account the unbalance of the random splitting data. The prediction performances of the models were evaluated by the accuracy and the area under the receiver operating characteristic curve (AUC).
Results: The AlexNet demonstrated the best prediction performance within a shortest calculation time of 151.1 s (accuracy = 90.9%, AUC = 0.974). GoogLeNet (accuracy = 89.1%, AUC = 0.978, calculation time = 223.5 s) and Inception_v3 (accuracy = 88.6%, AUC = 0.973, calculation time = 1402.3 s) showed worse performances than the AlexNet.
Conclusion: The DL-based models with transfer learning, especially AlexNet, could be feasible for the malignancy grade prediction of PGC using preoperative MR images.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI Grant Number JP17K15808.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
