Abstract:The study by Sui 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 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. However, despite the promising results, we are concerned about key aspects of the analysis. First, small (<1 cm) papillary thyroid cancers can be excluded given the increasing trend toward surveillance for this subgroup. With 2110 (22·9%) 0·3–1 cm clinically non-significant nodules assessed, the number of small nodules included in this study was large. Second, the malignancy rate (61%) was much higher than a previous study, which reported a malignancy rate of approximately 10%.2Durante C Grani G Lamartina L et al.The diagnosis and management of thyroid nodules: a review.JAMA. 2018; 319: 914-924Crossref PubMed Scopus (185) Google Scholar The reported pretest probability of malignancy is much higher than it will be if a deep-learning model is used in clinical practice. Consequently, the true specificity of ThyNet might be overestimated. It would be interesting to see how the deep-learning model functions with a validation set that has a pretest probability of malignancy of around 10%. Third, model construction might be too complex because the network architecture consists of three transfer learning methods (ResNet, ResNeXt, and DenseNet). This study did not prove the diagnostic accuracy of a combined architecture of three networks was higher than that of single network. Fourth, the diagnostic performance of radiologists without the assistance of ThyNet might have been underestimated by following strict diagnostic criteria and reviewing static ultrasound images. It would be of clinical significance to develop a deep-learning model based on dynamic videos instead of static images and then compare the diagnostic performance with radiologists in a real-world setting. A meta-analysis showed that experienced radiologists might have an advantage over deep-learning systems during real-time diagnosis.3Xu L Gao J Wang Q et al.Computer-aided diagnosis systems in diagnosing malignant thyroid nodules on ultrasonography: a systematic review and meta-analysis.Eur Thyroid J. 2020; 9: 186-193Crossref PubMed Scopus (12) Google Scholar Fifth, the test sets were collected from several centres that each used different ultrasound equipment and imaging parameters, it might be better to validate the deep-learning model in separate datasets with the same ultrasound machine. Sixth, there was a large difference in the experience of junior (1–3 years) and senior radiologists (>8 years) in thyroid imaging, but their diagnostic performance was very close, which confused us; the training protocols for junior and senior radiologists need clarification. Seventh, because thyroid cancer is a complex and heterogeneous disease, different types of medical data sources in practical clinical scenarios—such as patient demographics, laboratory findings (eg, thyroid hormone concentrations), and ultrasound images—should be combined to create a reliable diagnostic deep learning model, which was not done in this study. Finally, we feel that the classification performance of this study should be thoroughly tested in a different geographic setting. 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 managementIn The Lancet Digital Health, Sui Peng and colleagues,1 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,2 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,3 and if ThyNet can augment radiologists' diagnostic capabilities in general settings, it would be a breakthrough in managing thyroid nodules. Full-Text PDF Open Access

Deep learning diagnostic performance and visual insights in differentiating benign and malignant thyroid nodules on ultrasound images

Visual Interpretability in Computer-Assisted Diagnosis of Thyroid Nodules Using Ultrasound Images

Ultrasound Image Classification of Thyroid Nodules Based on Deep Learning

Deep learning for classification of thyroid nodules on ultrasound: validation on an independent dataset

Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic study

Artificial intelligence in thyroid ultrasound

Ultrasound-based deep learning using the VGGNet model for the differentiation of benign and malignant thyroid nodules: A meta-analysis

RF11 | PSAT234 Deep learning analysis of thyroid nodule ultrasound images has high sensitivity and negative predictive value to rule-out thyroid cancer

Deep learning approaches for differentiating thyroid nodules with calcification: a two-center study

Diagnostic performance of artificial intelligence in interpreting thyroid nodules on ultrasound images: a multicenter retrospective study

An Efficient Deep Convolutional Neural Network Model for Visual Localization and Automatic Diagnosis of Thyroid Nodules on Ultrasound Images.

A Novel Ultrasound Image Diagnostic Method for Thyroid Nodules

Deep learning for the diagnosis of suspicious thyroid nodules based on multimodal ultrasound images

Deep learning models for thyroid nodules diagnosis of fine-needle aspiration biopsy: a retrospective, prospective, multicentre study in China

The Auxiliary Diagnosis of Thyroid Echogenic Foci Based on a Deep Learning Segmentation Model: A Two-Center Study

Evaluation of a Deep Learning-Based Computer-Aided Diagnosis System for Distinguishing Benign from Malignant Thyroid Nodules in Ultrasound Images.

Research on The Identification of Benign and Malignant Thyroid Nodule Ultrasound Images Based on Deep Learning Model

An Integrated AI Model to Improve Diagnostic Accuracy of Ultrasound and Output Known Risk Features in Suspicious Thyroid Nodules

A Novel Ultrasound Image Diagnosis Method for Thyroid Nodules

A Deep-Learning Model to Assist Thyroid Nodule Diagnosis and Management

Thyroid Nodules Risk Stratification Through Deep Learning Based on Ultrasound Images