Using Artificial Intelligence To Diagnosis Thyroid Nodules Underlying Diffuse Thyroid Disease

Objective:Ultrasound is the first choice for diagnosing thyroid nodules,and diffuse thyroid disease can affect the sonographic characteristics of the nodules,making diagnosis more difficult,and more likely to cause misdiagnosis.Studies have shown that artificial intelligence performed well in diagnosing malignant nodules,but the ability of it to diagnose nodules underlying diffuse thyroid parenchyma has not been reported.Therefore,we aim to study the influence of diffuse thyroid parenchyma on the diagnosis ability of Ai,and to establish a thyroid nodule diagnosis model under diffuse thyroid disease.Methods:1428 cases with 4429 pictures of thyroid nodules from January 2017 to August 2018 were retrospectively collected and 80% were used as training set and 20% used as test set.452 cases with diffuse thyroid disease and nodules were collected diagnosed from September to December 2019,containing 1655 pictures.It was used as a diffusion test set.Patients with diffuse thyroid disease who have TI-RADS 4A-4C nodules were collected as difficult test set under diffuse thyroid disease.The set had 120 cases,569 pictures.We trained the RA Net,Res Net,and Dense Net models,evaluated their diagnostic results under the three test sets,and compared them with radiologist with different years of experience.From the above models,the best one was selected,and was constructed for two tasks: to identify diffuse thyroid parenchyma and to diagnose nodules in it.The model was optimized for diffuse thyroid disease.A total of 2932 cases with 9,106 pictures from January 2017 to August 2019 were collected based on the rule that patient with diffuse thyroid disease versus those without it should be equal to 1: 1.80% of the data were used as training set and 20%were used as test set.The ability of model to diagnose diffuse thyroid disease and nodules was evaluated on the test set.In addition,we compared the modified model with unmodified model on the diffuse test set and difficult test set under diffuse thyroid disease,and also compared with doctors.The statistical difference between AUC was calculated using Delong Test.Results:Under the normal test set,the AUC of RA Net,Res Net,and Dense Net was 0.864,0.908 and0.909,which were significantly higher than the doctors(p <0.001).Under the diffuse test set,the AUC of the models was 0.833,0.825 and 0.831,and AUC under the difficult test set was 0.684,0.743 and 0.726.All of them were not significantly different from those of the doctors(p> 0.05).The AUC for modified Dense Net under test set was 0.892 for diagnosing diffuse background and the AUC for nodule diagnosis was 0.914.Under the diffuse test set,the AUC of the modified model was 0.891,which was significantly higher than the original model and the doctors(p <0.01).Under the difficult test set,the AUC of the modified model was 0.755,which was significantly higher than that of the middle and low experienced doctors(p <0.05)and was similar with the original model and high experienced doctors(p>0.05).Conclusion:Diffuse thyroid disease can reduce the accuracy of deep learning models in diagnosing nodules.The modified model can effectively identify diffuse lesions and diagnose nodules.Its diagnostic ability is comparable to that of senior doctors,and better than senior doctors in some cases.