A multiple time-series deep learning (DL) model for computed tomography (CT) evaluation of ground-glass nodules (GGNs) offers significantly enhanced prognostic capability than clinical and radiomic models, according to a new study.
For the retrospective study, recently published in the European Journal of Radiology, researchers compared the aforementioned DL model, a clinical model, a delta radiomic model and a combined model, which incorporated DL, clinical features from electronic medical records (EMRs) and CT-derived semantic features. The total cohort was comprised of 486 patients (mean age of 56.7) and 1,031 CT scans, according to the study.
In the test set cohort, the study authors found that the combined model and deep learning model provided significantly higher sensitivity (84 percent and 88 percent respectively) for predicting malignant and benign GGNs in contrast to the delta radiomic model (68 percent) and clinical model (32 percent).
The clinical model and delta radiomic model offered higher specificity and positive predictive value (PPV) than the DL and combined models. The researchers noted a 95.8 percent specificity and an 88.9 percent PPV for the clinical model in comparison to 75 percent and 78.6 percent, respectively, for the DL model.
However, the study authors also noted that the DL model provided a 28.2 percent higher negative predictive value (NPV) than the clinical model (85.7 percent vs. 57.5 percent) as well as a 20.5 percent higher AUC (81.7 percent vs. 61.2 percent).
“ … Our DL model analyzes CT images at multiple time points, enabling a more comprehensive assessment of longitudinal changes in lesions and demonstrating satisfactory classification performance. … By combining … clinical information with (the) DL model, the model gains additional background information that aids in malignancy prediction, further enhancing its application value in predicting GGNs,” wrote lead study author Xiaolong Yang, M.D., who is affiliated with the Department of Radiology at Harbin Medical University in Heilongjiang, China, and colleagues.
Three Key Takeaways
- Superior sensitivity of DL models. The deep learning (DL) model analyzing multiple CT time points demonstrated higher sensitivity (88 percent) for predicting malignancy in ground-glass nodules (GGNs) than clinical (32 percent) and delta radiomic models (68 percent), enhancing early malignancy detection.
- Improved negative predictive value (NPV). The DL model yielded a significantly higher NPV (85.7 percent) and AUC (81.7 percent) compared to the clinical model (NPV: 57.5 percent, AUC: 61.2 percent), making it valuable in ruling out malignancy and reducing unnecessary follow-ups.
- Value of temporal and semantic CT features. By integrating longitudinal image changes and CT semantic features (e.g., lobulation, density heterogeneity), the model aids in distinguishing aggressive from indolent nodules, supporting more personalized surveillance strategies and minimizing over-treatment.
In differentiating malignant GGNs from benign GGNs, the researchers noted that malignant GGNs were characterized by a higher prevalence of part-solid nodules (59.8 percent vs. 36.6 percent), a higher percentage of nodules with heterogeneous density (71.3 percent vs. 47 percent) and a higher proportion of lobulation (81.9 percent vs. 31 percent).
“By capturing subtle temporal changes and incorporating CT semantic features like maximum diameter and lobulation, the model enhances predictive accuracy and helps distinguish between aggressive and indolent nodules,” maintained Yang and colleagues. “This approach not only reduces false positives but also minimizes unnecessary interventions, supporting a more personalized ‘watch-and-wait’ strategy and optimizing resource utilization in clinical practice.”
(Editor’s note: For related content, see “Can AI Predict Future Lung Cancer Risk from a Single CT Scan?,” “CT Study: Modified Lung-RADS Model Offers Enhanced Prognostic Assessment of Pure Ground-Glass Nodules” and “Can Radiomics Bolster Low-Dose CT Prognostic Assessment for High-Risk Lung Adenocarcinoma?”)
In regard to study limitations, the authors acknowledged the retrospective nature of the research and noted potential selection bias with a cohort entirely comprised of patients who had lung nodule resection.
