Optimized AI Assists Senior Radiologists in Thyroid Nodule Diagnostics

A study published in JAMA Network Open found that in the process of thyroid nodule management, an optimized artificial intelligence (AI) strategy assisted senior radiologists while a traditional AI strategy was a better fit for less experienced providers.

According to a study from the National Center for Biotechnology Information, there is a 0.1 percent estimated annual incidence of thyroid nodules in the US and a 10 percent lifetime chance of developing a thyroid nodule.

Although ultrasonography is effective for thyroid screening, radiologists are feeling workload pressure as the number of examinations increases along with interpretation time.

But AI could help in the diagnostic process, the researchers posited.

Continuing efforts to use this technology for thyroid nodule management, researchers aimed to establish the capabilities of optimized, personalized AI. Specifically, they aimed to determine whether an advanced solution can lessen the workload of radiologists while continuing high diagnostic performance.

In this study, researchers used a total of 1,754 preexisting ultrasonographic images of individual thyroid nodules from 1,048 patientsbetween July 1, 2018, and June 31, 2019. With this data, researchers constructed an optimized strategy that derived from how 16 junior and senior radiologists used AI-assisted diagnosis results with various image features.

Researchers also aimed to compare the optimized strategy with a traditional AI model. The data they used consisted of 300 ultrasonographic images from 268 patients with a total of 300 thyroid nodules from May 1, 2021, to December 31, 2021.

After reviewing diagnostic performances, time-based costs, and assisted diagnoses between the two models, researchers emerged with extensive data.

Of the 1,754 thyroid nodules that emerged from the retrospective dataset, the mean size was 16.4 mm and 42.6 percent were harmless while 57.4 percent were dangerous.

Of the 300 thyroid nodules derived from the prospective dataset, the mean size was 17.2 mm and 41.7 percent were harmless while 58.3 percent were dangerous.

Researchers found that the effects of the optimized AI strategy were not similar between junior and senior radiologists. Particularly, the two groups of radiologists liked the optimized AI model better for different types of nodules.

For example, junior radiologists did not prefer using the optimized AI for cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, and nodules smaller than 5 mm.

Meanwhile, cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, very hypoechoic nodules, nodules that were taller rather than wider, lobulated or irregular nodules, and extrathyroidal extension were among the features that did not benefit from advanced AI among senior radiologists.

However, junior radiologists also experienced slower mean task completion times when using the optimized strategy. Contrarily, this new method led to faster completion times for senior radiologists.

For example, when using the optimized AI strategy, 2 separate junior radiologists experienced completion times that were 4.2 and 2.9 seconds slower. Meanwhile, 2 separate senior radiologists experienced completion times that were 2.6 and 2.5 seconds faster. 

Based on these findings, researchers emerged with the conclusion that an optimized AI strategy may be a better fit for senior radiologists. This was mainly because of its ability to shorten diagnostic time-based costs while maintaining diagnostic accuracy for these providers.

New research continues to show that the healthcare industry is discovering ways to leverage AI to mediate staffing shortages.

A March survey from the Health Management Academy found that many C-suite executives were using AI for clinical procedures due to common staffing predicaments.

Given the growing issue of limited staffing, the use of AI is becoming increasingly common. Specifically, the survey found that the share of executives using AI for workforce issues was 47.5 percent, while the remaining 52.5 percent were considering AI for this purpose.