Deep Learning For Automatic Segmentation And Diagnosis Of Parotid Tumors

To establish a deep learning model for fully automatic segmentation and diagnosis of benign and malignant parotid tumors. In total, 493 patients with pathologically confirmed parotid tumors (396 benign, 97 malignant) were included. Among them, images from 82 patients (75 benign, 7 malignant) scanned on different MRI machines served as an external validation set. Firstly, an automatic segmentation model based on nnU-NetV2 was developed for the tumor segmentation. Segmentation performance was evaluated by DICE coefficient. Secondly, based on the results of automatic segmentation, a Resnet-18 deep learning model was constructed for the classification of benign and malignant parotid tumors. Finally, a combined diagnostic model was built via logistic regression of the deep learning, radiomics, and clinical models. The automatic segmentation model showed a DICE coefficient of 0.86 and 0.82 in the internal and external validation sets, respectively. The combined diagnostic model, which was trained using the automatically segmented fat-suppressed T2-weighted imaging, demonstrated an area under the curve (AUC) of 0.92 (95% confidence interval (CI): 0.86–0.98) and 0.90 (95% CI: 0.77–1.00) in the internal and external validation sets, respectively, with corresponding accuracy rates of 91.1% and 93.9%. An MRI-based deep learning model can achieve fully automatic segmentation and classification of benign and malignant parotid tumors.