Radiologists Versus AI-Based Software: Predicting Lymph Node Metastasis and Prognosis in Lung Adenocarcinoma From CT Under Various Image Display Conditions.

PURPOSE

To compare the variability of quantitative values from lung adenocarcinoma CT images independently assessed by 2 radiologists and AI-based software under different display conditions, and to identify predictors of pathological lymph node metastasis (LNM), disease-free survival (DFS), and overall survival (OS).

METHODS

Preoperative CT images of 307 patients were displayed under 4 conditions: lung-1, lung-2, mediastinum-1, and mediastinum-2. Two radiologists (R1, R2) measured total diameter (tD) and the longest solid diameter (sD) under each condition. The AI-based software automatically detected lung nodules, providing tD, sD, total volume (tV), and solid volume (sV).

RESULTS

All measurements by R1 and R2 with AI-based software were identical. Four out of the 8 measurements showed significant variation between R1 and R2. For LNM, multivariate logistic regression identified significant indicators including sD at mediastinum-2 of R1, sD at mediastinum-1 and mediastinum-2 of R2, tV, and the proportion of sV to tV (sV/tV) of AI-based software. For DFS, multivariate Cox regression identified sD at lung-1 of R1, the proportions of sD to tD at lung-2 of R1, sD at lung-2 and mediastinum-1 of R2, tV, and sV/tV of AI-based software as significant. For OS, multivariate Cox regression identified sD at lung-1 and mediastinum-2 of R1, tD at lung-2 of R2, sD at mediastinum-1 of R2, sV, and sV/tV of AI-based software as significant.

CONCLUSION

Radiologists' CT measurements were significant predictors of LNM and prognosis, but variability existed among radiologists and display conditions. AI-based software can provide accurate and reproducible indicators for predicting LNM and prognosis.