Deep Learning-based Computerized Tomographic Imaging for Differentiation and Segmentation of Parotid Gland Neoplasm

We applied a convolution neural network (CNN) to the parotid tumor classification and segmentation. The bounding box prediction of CNNs was used to detect the areas of parotid tumors. The Yolov4 method was used to obtain AP₅₀ 0.964. Furthermore, the ResNet+CBAM and ResNet+BiFPN were applied to classify each image into mixed, Warthin, and malignant tumors. The classification accuracies of ResNet+BiFPN and ResNet+CBAM were 0.8526 and 0.8419 (for mixed malignant and Warthin) and 0.8216 and 0.8111 (for mixed malignant). To effectively classify the slice images of patients and normal participants, we developed a decision tree to integrate classified images to make a decision. Using the U-net and Unet ++, we segmented the tumors of images. For 1493 tumor images, the performances of U-net and Unet ++ were presented as the Dice measure of 0.850 and 0.863. The results revealed that the classification and the segmentation showed an accuracy of 87% and a Dice coefficient of 0.91.