Bing Yu,1,2,* Huijuan He,3,* Qiao Zheng,2,* Yao Ai,2 Xianwen Yu,2 Sunjun Li,4 Ji Zhang,2 Juebin Jin,5 Xiance Jin,2,6 Wenliang Yu3 

1Purchasing Center Department, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, 310020, People’s Republic of China; 2Radiation and Medical Oncology Department, Wenzhou Medical University First Affiliated Hospital, Wenzhou, 325000, People’s Republic of China; 3Radiation Oncology Department, Quzhou People’s Hospital, Quzhou, 324002, People’s Republic of China; 4Alberta College, Wenzhou Medical University, Wenzhou, 325000, People’s Republic of China; 5Medical Engineering Department, Wenzhou Medical University First Affiliated Hospital, Wenzhou, 325000, People’s Republic of China; 6School of Basic Medical Science, Wenzhou Medical University, Wenzhou, 325000, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Xiance Jin, Radiation and Medical Oncology Department, Wenzhou Medical University First Affiliated Hospital, Wenzhou, 325000, People’s Republic of China, Email jinxc1979@hotmail.com Wenliang Yu, Radiation Oncology Department, Quzhou People’s Hospital, Quzhou, 324002, People’s Republic of China, Email wlyu1982@aliyun.com

Purpose: The feasibility and accuracy of ultrasound-based radiomics, deep learning, and combined deep learning radiomics models were investigated in the differentiation of papillary thyroid carcinoma and papillary thyroid microcarcinoma to decrease the risk of overtreatment of papillary thyroid microcarcinoma.
Methods: A total of 549 patients with confirmed 180 papillary thyroid carcinoma and 436 papillary thyroid microcarcinoma nodules from Hospital One were enrolled and randomly divided into training and validation cohorts at a ratio of 8:2 with 56 patients left as independent testing set 1. Fifty patients from Hospital Two were enrolled as independent testing set 2. Radiomics signature and five deep learning networks, such as visual geometry group 13 (VGG13), VGG16, VGG19, AlexNet, and EfficientNet, were generated for papillary thyroid carcinoma and papillary thyroid microcarcinoma differentiation. Combined deep learning and radiomics models were constructed to further improve the differentiation ability.
Results: An area under curves of 0.826 and 0.822 was achieved with radiomics model for papillary thyroid carcinoma and papillary thyroid microcarcinoma differentiation in the independent testing set 1 and set 2, respectively. VGG19 achieved the best area under curves of 0.890 and EfficientNet achieved the best accuracy of 0.867. The best accuracy and area under curves of 0.904, 0.900, and 0.931, 0.946 were achieved with the combination of VGG + radiomics (R_V_Combined) and EffiecientNet + radiomics (R_E_Combined) in the independent testing set 1 and set 2, respectively.
Conclusion: Deep learning and radiomics combination models are promising in the noninvasively preoperative differentiation of papillary thyroid microcarcinoma and papillary thyroid carcinoma to decrease the overtreatment of patients with papillary thyroid microcarcinoma and to minimize the complications caused by overtreatment.

Keywords: papillary thyroid carcinoma, papillary thyroid microcarcinoma, ultrasound, deep learning, radiomics