The feasibility study of using deep learning to identify the pathological subtypes of parotid gland tumors based on MR images

WANG Yan; TIAN Hui; HONG Yue; LI Hui

Abstract：Objective To automatically identify pathological types of parotid gland tumors using deep learning methods based on multiple parameters of MR T2,DWI and enhanced T1 images,and to evaluate the efficacy of deep learning by compar-ing with pathological results.Methods MRI images of 39 cases of parotid gland tumors confirmed by pathology,including 9 cases of pleomorphic adenoma,13 cases of Warthin tumor,6 cases of malignant tumor,and 11 cases of other non-tumor lesions,were selected in this work.After MR images were standardized,2D U-NET was used to identify tumor types,and the classifica-tion efficiency through the input of single-channel of T2 image,and multi-channel of T2,DWI and enhanced T1 image were com-pared respectively.The data of 29 cases were used for training sets and the data of 10 cases for test sets.Accuracy,sensitivity,specificity,F1 score and precision were calculated to evaluate the efficacy of tumor classification.Results When training with T2 single-channel input,Warthin tumor had the highest F1 score(59.2%),sensitivity(51.1%)and precision(70.3%),followed by pleomorphic adenoma.The sensitivity,accuracy and F1 scores of multi-channel input for Warthin tumor identification were 79.5%,70.0%and 74.5%,respectively,the sensitivity and specificity of polymorphic adenoma were 46.2%and 81.5%,respec-tively.Conclusion By analyzing the imaging characteristics of parotid gland tumors,the deep learning method can effectively identify Warthin tumor,and the combination of multi-parametric MR images can improve the sensitivity of Warthin tumor and pleomorphic adenoma recognition.