Room: Track 1
Purpose: resonance imaging (MRI) is majority of imaging modality to differentiate benign from malignant parotid gland cancer, which is clinically important. This study aims to develop and evaluate a deep learning network for distinction between benign and malignant parotid gland tumors via learning MR images.
Methods: whole dataset consisting of two-hundred-forty-nine patients with parotid gland tumors. All patient’s MR images crop 2396 parotid glands and tumors region images in total. Each image assignment a label (free of tumor, benign tumor and malignant tumor) by histology results. These data were randomly divided into 90% training dataset and 10% test dataset. Training dataset including 2176 images from 224 patients (752 free of tumor, 538 malignant tumors, and 886 benign tumors) and 220 images from 25 patients in test data (76 free of tumor, 74 malignant tumors, and 70 benign tumors). In order to achieve effective classification results, the res-net model was modified. The input images include 3 channels (T1 or T2 parotid gland, T1 tumor only, T2 tumor only), 224x224 pixels in each channel. All images underwent randomly data enhancements by image flipping, contrast adjustment. The learning rate of 1e-6, after 1500 epoch training, the model gradually converges. Develop entire program with Pytorch (version 1.2).
Results: model sensitivity and specificity for benign and malignant discrimination were 84.01% (95%CI [0.80, 0.87]), 92.45% (95%CI [0.90, 0.94]) for training dataset, and 78.57% (95%CI [0.67, 0.87]),85.71% (95%CI [0.75, 0.93]) for test dataset. The accuracy of model was 89% (95%CI [0.87, 0.90]) and 83% (95%CI [0.79, 0.88]) for training and test dataset. The mirco-AUC of model was 0.84 for test dataset and 0.96 for training dataset.
Conclusion: this study propose model expected to drastically enhance clinicians effective in diagnosis of parotid tumors. Nonetheless, large-scale multicenter studies are required for a full validation in the future.
IM/TH- Image Analysis Skills (broad expertise across imaging modalities): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)