Predicting The Benignancy Or Malignancy Of Parotid Tumor Using Deep Learning

Objective:Evaluating the diagnostic performance of four deep learning models for discriminating benignancy or malignancy in Parotid tumor based on enhanced CT images.Methods:From January 2019 to January 2023,clinical and image data of 165 patients with Parotid tumor admitted to Shanxi Provincial People’s Hospital were collected.Based on surgical pathology results,there were 131 cases of benign tumors and 34 cases of malignant tumors.The data was randomly divided into training and test sets in a 7:3 ratio,and 1325 enhanced CT images of parotid tumors were extracted.The training set included 928 images from 116 patients,and the test set included 397 images from 49 patients.Four deep learning models(Res Net50,Dense Net121,Efficient Net B3,Vi T-B\16)were trained on the training set and used to classify the benign or malignant tumors in the testing set.Additionally,two radiologists with different levels of experience also independently classified the raw enhanced CT images of the 49 patients in the testing set into benign or malignant categories.The diagnostic performance accuracy,sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV),F1 score,and area under the curve(AUC))were calculated for each of the four deep learning models and the two radiologists,and their diagnostic performance was compared.Results:In diagnosing the benign or malignant tumors in the test set of parotid,the sensitivity and specificity of Res Net50,Dense Net121,Efficient Net B3,Vi T-B\16 were 79.27% and77.78%,76.83% and 76.83%,79.27% and 77.14%,80.49% and 79.68%,respectively.Regarding the classification results of parotid tumors for the 49 patients in the test set,the sensitivity,specificity,and AUC of Res Net50,Dense Net121,Efficient Net B3,Vi T-B\16,Junior and senior radiologists were 80.00%,79.49%,and 0.80;70.00%,74.36%,and 0.72;80.00%,82.05%,and 0.81;90.00%,84.62%,and 0.87;60.00%,71.79%,and 0.66;100%,89.74%,and 0.95.Conclusion:1.Four deep learning models have good diagnostic efficacy in predicting benign and malignant parotid tumors on enhanced CT images.The models with the highest to lowest diagnostic efficacy are Vi T-B\16,Efficient Net B3,Res Net50,and Dense Net121.2.In predicting benign and malignant parotid tumors on enhanced CT images,although the diagnostic efficacy of the four deep learning models was worse than that of the senior radiologists,the diagnostic time of the models was significantly less than that of the radiologists.The more radiologists use deep learning models in their daily tasks,the more time they will have to focus on intellectually demanding tasks,and in the future deep learning may replace radiologists’ image interpretation capabilities..