A deep-learning model to assist thyroid nodule diagnosis and management

In The Lancet Digital Health, Sui Peng and colleagues,1Peng S Liu Y Lv W et al.Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic study.Lancet Digit Health. 2021; 3: e250-e259Summary Full Text Full Text PDF PubMed Scopus (7) Google Scholar showed the ability of a deep-learning model (ThyNet) to improve radiologists' diagnostic performance when determining the nature of a thyroid nodule. Unlike previous studies that compared the accuracy of humans and artificial intelligence (AI) performance,2Buda M Wildman-Tobriner B Hoang JK et al.Management of thyroid nodules seen on US images: deep learning may match performance of radiologists.Radiology. 2019; 292: 695-701Crossref PubMed Scopus (51) Google Scholar this study is important because it shows that the future of diagnostics is cooperation and synergy between humans and AI, rather than exclusivity and usurpation. Radiological diagnosis of thyroid nodules can be difficult in many situations, even with the American College of Rheumatology Thyroid Imaging Reporting and Data System,3Tessler FN Middleton WD Grant EG et al.ACR thyroid imaging, reporting and data system (TI-RADS): white paper of the ACR TI-RADS committee.J Am Coll Radiol. 2017; 14: 587-595Summary Full Text Full Text PDF PubMed Scopus (689) Google Scholar and if ThyNet can augment radiologists' diagnostic capabilities in general settings, it would be a breakthrough in managing thyroid nodules. However, there are issues that should be addressed. Peng and colleagues1Peng S Liu Y Lv W et al.Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic study.Lancet Digit Health. 2021; 3: e250-e259Summary Full Text Full Text PDF PubMed Scopus (7) Google Scholar report that 8339 patients were included in the training set, 1424 in test set A, 1048 in test set B, and 303 in test set C. However, the appendix indicates that all patients were diagnosed either with papillary, follicular, medullary, or anaplastic cancers, which implies that all learning and testing materials consisted only of malignancies. Moreover, figure 1 states that 18 049 images from 5122 malignant and 3217 benign pathologically proven nodules were used in the training set, indicating that more than one image was obtained from a nodule, but the selection process of images from each nodule has not been clarified. Furthermore, the precise numbers of malignant and benign nodules in test sets A, B, and C seem to be missing. Likewise, the number of benign and malignant nodules that were diagnosed correctly and incorrectly by the radiologists without ThyNet, ThyNet-only, and the radiologists with ThyNet should also be presented along with the sensitivity, specificity, and positive and negative predictive values. ThyNet's diagnostic ability to identify follicular thyroid carcinomas should be properly evaluated in accordance with actual clinical settings, in which, many types of benign (nodular hyperplasia, follicular adenoma, and hurthle cell adenoma) and malignant (follicular thyroid carcinomas, Hurthle cell carcinoma, and follicular variant papillary thyroid carcinoma) nodules with follicular patterns that are difficult to differentiate. Proportioning the number of benign and malignant subtypes in the training set and test set A, B, and C to be representative of the real clinical setting would empower ThyNet with greater clinical relevance.4Chai YJ Song J Shaear M Yi KHJAoT Artificial intelligence for thyroid nodule ultrasound image analysis.Ann Thyroid. 2020; 5: 8Crossref Google Scholar Likewise, presenting ThyNet's performance results in diagnosing each pathological subtypes from the whole group would help readers to comprehensively understand the ability of the AI. We declare no competing interests. Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic studyThe ThyNet-assisted strategy can significantly improve the diagnostic performance of radiologists and help reduce unnecessary fine needle aspirations for thyroid nodules. Full-Text PDF Open AccessA deep-learning model to assist thyroid nodule diagnosis and management – Authors' replyWe thank Joon-Hyop Lee and Young Jun Chai for their insightful comments about our work1 and important suggestions for improvements to our analyses. Full-Text PDF Open AccessA deep-learning model to assist thyroid nodule diagnosis and managementThe study by Sui Peng and colleagues1 published in The Lancet Digital Health adds to the growing body of evidence that the application of deep-learning (or ensemble transfer learning) models can significantly improve the diagnostic performance of both junior and senior radiologists and reduce the number of unnecessary biopsies. Full-Text PDF Open Access