The Value Of S-Detect In Improving The Diagnostic Performance Of Radiologists For The Differential Diagnosis Of Thyroid Nodules

Qi Wei,Shu-E Zeng,Li-Ping Wang,Yu-Jing Yan,Ting Wang,Jian-Wei Xu,Meng-Yi Zhang,Wen-Zhi Lv,Xin-Wu Cui,Christoph F Dietrich
DOI: https://doi.org/10.11152/mu-2501
2020-01-01
Medical Ultrasonography
Abstract:Aims: To compare the diagnostic value of S-Detect (a computer aided diagnosis system using deep learning) un differentiating thyroid nodules in radiologists with different experience and to assess if S-Detect can improve the diagnostic performance of radiologists. Materials and methods: Between February 2018 and October 2019, 204 thyroid nodules in 181 patients were included. An experienced radiologist performed ultrasound for thyroid nodules and obtained the result of S-Detect. Four radiologists with different experience on thyroid ultrasound (Radiologist 1, 2, 3, 4 with 1, 4, 9, 20 years, respectively) analyzed the conventional ultrasound images of each thyroid nodule and made a diagnosis of "benign" or "malignant" based on the TI-RADS category. After referring to S-Detect results, they re-evaluated the diagnoses. The diagnostic performance of radiologists was analyzed before and after referring to the results of S-Detect. Results: The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of S-Detect were 77.0, 91.3, 65.2, 68.3 and 90.1%, respectively. In comparison with the less experienced radiologists (radiologist 1 and 2), S-Detect had a higher area under receiver operating characteristic curve (AUC), accuracy and specificity (p <0.05). In comparison with the most experienced radiologist, the diagnostic accuracy and AUC were lower (p<0.05), In the less experienced radiologists, the diagnostic accuracy, specificity and AUC were significantly improved when combined with S-Detect (p<0.05), but not for experienced radiologists (radiologist 3 and 4) (p>0.05). Conclusions: S-Detect may become an additional diagnostic method for the diagnosis of thyroid nodules and improve the diagnostic performance of less experienced radiologists.
What problem does this paper attempt to address?