深度学习预测非小细胞肺癌脏层胸膜侵犯

BACKGROUND: To evaluate and compare the diagnostic performance of several commonly used deep learning (DL) models and a conventional clinical-radiological feature model based on preoperative computed tomography (CT) for predicting visceral pleural invasion (VPI) in non-small cell lung cancer (NSCLC).

METHODS: A retrospective cohort of 3,406 patients from four hospitals was analyzed. Data were split into training, validation, and test sets. A logistic regression model using clinical-radiological features was compared against five DL architectures (including Res2Net_3D8F and ResNet variants). Performance was evaluated via AUC, sensitivity, specificity, PPV, and NPV.

RESULTS: In the independent test set, the clinical-radiological feature model achieved the highest AUC (0.875), but with low sensitivity (0.21) and high specificity (0.96). The best-performing DL model (Res2Net_3D8F) achieved an AUC of 0.835, with sensitivity of 0.83, specificity of 0.74, PPV of 0.36, and NPV of 0.96. All DL models demonstrated higher sensitivity (0.71-0.83) but lower PPV (0.29-0.36) compared to the clinical model. There was no statistically significant difference in AUC between the Res2Net_3D8F model and the clinical-radiological model in the test set (P = .123, DeLong test).

CONCLUSIONS: Although both clinical and DL models based on CT imaging showed moderate discriminatory ability for preoperative VPI prediction, neither approach achieved optimal clinical utility. The clinical model had high specificity but very low sensitivity, while DL models had improved sensitivity but low PPV and did not outperform the clinical model. Noninvasive VPI prediction remains a significant challenge, highlighting the need for further methodological advancements and validation.