Comparing an emerging computed tomography (CT)-based radiomics model to conventional CT in differentiating between benign and malignant thyroid nodules, the authors of a new study found that the CT radiomics model offered a nearly 10 percent higher area under the curve (AUC), 13 percent higher accuracy and nearly 27 percent higher specificity.
For the retrospective study, recently published in the American Journal of Roentgenology, researchers assessed the capability of a CT-based radiomics model that combines patient age, three key morphologic features (including abnormal cervical lymph nodes) and 28 radiomic features drawn from non-contrast, arterial and venous phases of neck CT exams. The cohort was comprised of 378 patients (mean age of 46.3) who had a total of 408 resected thyroid nodules, according to the study.
In comparison to conventional CT, external validation testing revealed that the CT radiomics model offered a significantly higher AUC (92.3 percent vs. 82.7 percent), sensitivity (84 percent vs. 76.5 percent), specificity (94.1 percent vs. 67.6 percent), and accuracy (87 percent vs. 73.9 percent) for the differentiation of benign and malignant thyroid nodules.
“We do not anticipate that CT radiomics models would replace ultrasound interpreted by ACR T1-RADS in the evaluation of thyroid nodules. However, CT radiomics models could provide a useful complement to current tools, for example as a troubleshooting option for nodules that remain challenging to characterize by ultrasound. Additionally, the radiomics models could be useful to guide further decision-making for thyroid nodules detected incidentally by CT,” wrote study co-author Daiying Lin, M.D., who is associated with the Department of Radiology at Shantou Central Hospital in Shantou, China, and colleagues.
In the training and internal validation sets for the CT radiomics model, the researchers noted interreader agreement for the two reviewing radiologists of 86 percent for cystic change, 85 percent for the edge interruption sign and 83 percent for abnormal cervical lymph nodes.
Specifically, the study authors noted that in the training set for the CT-based radiomics model, malignant nodules had a 44.3 percent lower frequency of cystic changes (3.2 percent vs, 47.5 percent) and a 43.6 percent higher frequency of the edge interruption sign (52.4 percent vs. 8.8 percent).
“(Cystic change) is more common in benign nodules, possibly due to development of surrounding fibrosis and fibrous membranes that affect these nodules’ blood supply, leading to hemorrhage, necrosis, and cystic areas,” pointed out Lin and colleagues.
“In contrast, the edge interruption sign indicates possible involvement of the thyroid capsule by a malignant nodule in proximity to the capsule. An earlier study found this finding to be useful for identification of papillary thyroid microcarcinomas.”
Three Key Takeaways
- Improved diagnostic performance. The CT radiomics model demonstrated superior diagnostic performance compared to conventional CT in distinguishing between benign and malignant thyroid nodules. With nearly 10% higher AUC, 13% higher accuracy, and almost 27% higher specificity, the CRT radiomics model presents a promising tool for enhancing diagnostic accuracy in clinical practice.
- Quantified impact of morphologic features. The study emphasizes the tangible impact of morphologic features, such as the 44.3% lower frequency of cystic changes and the 43.6% higher frequency of the edge interruption sign associated with malignant nodules compared to benign nodules. These statistically significant differences highlight the pivotal role of morphologic characteristics in augmenting the diagnostic precision of the CT radiomics model.
- Complementary tool in clinical practice. While not intended to replace ultrasound-based evaluations, CT radiomic models offer a complementary approach, especially for challenging cases in which ultrasound characterization remains ambiguous. Additionally, they can aid in decision-making for incidentally detected thyroid nodules on CT scans, potentially guiding further management strategies for these incidental findings.
(Editor’s note: For related content, see “What New Research Reveals About ChatGPT and Ultrasound Detection of Thyroid Nodules,” “CT Update: FDA Changes Course on Post-ICM Thyroid Monitoring in Young Children” and “Pediatric Thyroid Nodules on Ultrasound: Deep Learning Model and TI-RADS Show Higher Sensitivity than Radiologist Assessment.”)
The researchers also suggested that the absence of abnormal cervical lymph nodes as a feature in prior morphologic CT models may have been a factor in the lower performance of these models in differentiating thyroid nodules.
In the training set for their CT-based radiomics model, the study authors found that malignant thyroid nodules were associated with a 24.8 percent higher frequency of abnormal cervical lymph nodes in comparison to benign thyroid nodules (37.3 percent vs. 12.5 percent).
In regard to study limitations, the authors conceded that the high fraction of malignant nodules in the study sets may be due to the study’s focus on surgically resected nodules. The researchers noted that manual segmentation of images may have increased the variability with extracted features. While some radiomics features were related to nodule size, the study authors acknowledged that these features were not included in radiomic models.
