Optimized AI Strategy May Improve Thyroid Nodule Diagnosis and Management

An optimized integration of artificial intelligence (AI) decision aids in thyroid nodule management may help more experienced radiologists reduce diagnostic time-based cost without sacrificing diagnostic accuracy, whereas, a traditional all-AI strategy had preferable time-based cost for less experienced radiologists, according to study findings in JAMA Network Open. 

Integrating AI and digital imaging decision aids into medical diagnostics has improved ultrasonographic identification of thyroid nodules. Some studies have shown the diagnostic capability of AI was comparable to that of radiologists. More recently, researchers conducted a diagnostic study to develop AI decisions to help reduce the workload of radiologists. In this study, the use of an optimized diagnostic aid was compared with that of a traditional all-AI strategy among radiologists with varying levels of expertise. 

The researchers collected 2 groups of images: retrospective and prospective. A retrospective set of ultrasonographic images was used to optimize AI diagnosis strategy, and a prospective set of images was then used to compare the diagnostic performance and cost between the optimized AI strategy and the traditional strategy. The use of AI by 16 radiologists was assessed and significant and insignificant AI-assisted features were identified.

A total of 1754 ultrasonographic images of 1048 patients were collected retrospectively. The participants included 299 men and 749 women aged an average of 42.1 years (standard deviation [SD], 13.2 years) with a total of 748 (42.6%) benign nodules and 1006 (57.4%) malignant thyroid nodules, with an average size of 10.6 mm. For the prospective set, the researchers collected 300 ultrasonographic images of 268 patients; images from 74 men and 194 women aged an average of 41.7 years (SD, 14.1 years) were included. Among the prospective group there were 125 (41.7%) benign and 175 (58.3%) malignant thyroid nodules, with an average size of 17.2 mm. The participants included in the study were adults aged 18 years and older with confirmed benign or malignant thyroid nodules.

Among radiologists with less experience, the sonographic features that were not significantly improved by AI were cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, and nodules smaller than 5 mm. For more experienced radiologists, the insignificant features were cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, very hypoechoic nodules, nodules taller than wide, lobulated or irregular nodules, and extrathyroidal extension.

The researchers found radiologists with less experience reported a slight decrease in specificity when using AI assistance in identifying cystic or almost completely cystic nodules (from 98% to 97%) anechoic nodules (from 98% to 97%), and nodule size less than 5 mm (from 68% to 67%).

Compared with traditional AI strategy, optimized AI use was associated with increased mean task completion times for radiologists with less experience, increasing from 15.2 seconds (95% CI, 13.2-17.2 seconds) to 19.4 seconds (95% CI, 15.6-23.3 seconds) in one analysis, and from 12.7 seconds (95% CI, 11.4-13.9 seconds) to 15.6 seconds (95% CI, 13.6-17.7 seconds) in another. 

Nonetheless, the optimized strategy allowed for shorter task completion times for more experienced radiologists decreasing time from 19.4 seconds (95% CI, 18.1-20.7 seconds) to 16.8 seconds (95% CI, 15.3-18.3 seconds) in one set and from 12.5 seconds (95% CI, 12.1-12.9 seconds) to 10.0 seconds (95% CI, 9.5-10.5 seconds).

Study limitations include the lack of a crossover design, not exploring deep learning interpretability of algorithms, and conducting the study in a clinical environment only. 

“As our results show, the traditional all-AI strategy was preferable for junior radiologists, whereas the optimized strategy was better suited for senior radiologists,” the researchers concluded. “These results reveal a promising approach to implementing AI-assisted precision medicine.”