JMIR Publications

ABSTRACT

Background:

Early and accurate detection of thyroid cancer is crucial for effective treatment. This study evaluates the diagnostic performance of deep learning (DL) algorithms in identifying thyroid nodules (TN).

Objective:

To assess the diagnostic accuracy of DL algorithms in detecting and segmenting thyroid nodules through a meta-analysis.

Methods:

A total of 41 studies were included, with 14 focused on segmentation and 27 on detection tasks. Subgroup analyses were conducted based on the use of transfer learning (TL), application of algorithms, and clinician involvement.

Results:

For segmentation tasks, the pooled sensitivity (SE) was 82% (95% CI 79-84%) and specificity (SP) was 95% (95% CI 92-96%), with an area under the curve (AUC) of 0.91 (95% CI 0.89-0.94). For detection tasks, the pooled SE was 91% (95% CI 89-93%) and SP was 89% (95% CI 86-91%), with an AUC of 0.96 (95% CI 0.93-0.97).

Conclusions:

DL algorithms show promising potential in detecting TNs, with performance comparable to that of human clinicians. However, the studies included were often poorly designed and reported, which may lead to biased and overestimated results. Future research should adopt standardized methodologies and reporting guidelines to enhance the quality of DL studies. Clinical Trial: The study was registered with the PROSPERO international register of systematic reviews under the number CRD42024599495.