Classification of Parotid Gland Neoplasms in Computed Tomography Images Using Convolutional Neural Networks

This study investigates the application of convolutional neural networks (CNNs) for the classification and segmentation of parotid tumors. The classification process involves two stages: first, isolating the region of the parotid gland, and second, classifying it into one of three categories—mixed, malignant, or Warthin tumors. The Yolov7 method achieved an AP50 of 0.964 for the initial stage. Additionally, the data-efficient image transformers (DEIT) method was employed to classify parotid gland computerized tomographic images into mixed, Warthin, or malignant tumor categories. The classification accuracies were 0.923 for distinguishing between tumor presence and absence, 0.947 for differentiating malignant tumors from mixed and Warthin tumors, and 0.844 for distinguishing malignant from mixed tumors. To enhance diagnostic accuracy across all computerized tomographic slices for each patient or healthy participant, a decision tree mechanism based on the DEIT model was developed, aggregating classification results from individual slices. Furthermore, U-Net, U-Net++, TransUNet, and Swin-Unet models were independently applied to segment tumor images. Among 2,961 tumor images tested, TransUNet demonstrated superior performance, achieving an average Dice similarity coefficient of 0.921. Experimental results indicate that the combined classification and segmentation approach achieves an overall classification accuracy of 0.904 and a Dice similarity coefficient of 0.921.