A deep learning model for the differential diagnosis of benign and malignant salivary gland tumors based on ultrasound imaging and clinical data.

Abstract

The preoperative differentiation between benign parotid gland tumors (BPGTs) and malignant parotid gland tumors (MPGTs) is of great significance for therapeutic decision-making. Deep learning (DL), an artificial intelligence algorithm based on neural networks, can help overcome inconsistencies in conventional ultrasonic (CUS) examination outcomes. Therefore, as an auxiliary diagnostic tool, DL can support accurate diagnosis using massive ultrasonic (US) images. This current study developed and validated a DL-based US diagnosis for the preoperative differentiation of BPGT from MPGT. A total of 266 patients, including 178 patients with BPGT and 88 patients with MPGT, were consecutively identified from a pathology database and enrolled in this study. Ultimately, considering the limitations of the DL model, 173 patients were selected from the 266 patients and divided into 2 groups: a training set, and a testing set. US images of the 173 patients were used to construct the training set (including 66 benign and 66 malignant PGTs) and testing set (consisting of 21 benign and 20 malignant PGTs). These were then preprocessed by normalizing the grayscale of each image and reducing noise. Processed images were imported into the DL model, which was then trained to predict the images from the testing set and evaluated for performance. Based on the training and validation datasets, the diagnostic performance of the 3 models was assessed and verified using receiver operating characteristic (ROC) curves. Ultimately, before and after combining the clinical data, we compared the area under the curve (AUC) and diagnostic accuracy of the DL model with the opinions of trained radiologists to evaluate the application value of the DL model in US diagnosis. The DL model showed a significantly higher AUC value compared to doctor 1 + clinical data, doctor 2 + clinical data, and doctor 3 + clinical data (AUC =0.9583 vs. 0.6250, 0.7250, and 0.8025 respectively; all P<0.05). In addition, the sensitivity of the DL model was higher than the sensitivities of the doctors combined with clinical data (97.2% vs. 65%, 80%, and 90% for doctor 1 + clinical data, doctor 2 + clinical data, and doctor 3 + clinical data, respectively; all P<0.05). The DL-based US imaging diagnostic model has excellent performance in differentiating BPGT from MPGT, supporting its value as a diagnostic tool for the clinical decision-making process.