RATIONALE AND OBJECTIVES

To develop and validate a deep learning model for automated pathological grading and prognostic assessment of lung cancer using CT imaging, thereby providing surgeons with a non-invasive tool to guide surgical planning.

MATERIAL AND METHODS

This study utilized 572 cases from the National Lung Screening Trial cohort, dividing them randomly into training (461 cases) and internal validation (111 cases) sets in an 8:2 ratio. Additionally, 224 cases from four cohorts obtained from the Cancer Imaging Archive, all diagnosed with non-small cell lung cancer, were included for external validation. The deep learning model, built on the MobileNetV3 architecture, was assessed in both internal and external validation sets using metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). The model's prognostic value was further analyzed using Cox proportional hazards models.

RESULTS

The model achieved high accuracy, sensitivity, specificity, and AUC in the internal validation set (accuracy: 0.888, macro AUC: 0.968, macro sensitivity: 0.798, macro specificity: 0.956). External validation demonstrated comparable performance (accuracy: 0.807, macro AUC: 0.920, macro sensitivity: 0.799, macro specificity: 0.896). The model's predicted signatures correlated significantly with patient mortality and provided valuable insights for prognostic assessment (adjusted HR 2.016 [95% CI: 1.010, 4.022]).

CONCLUSIONS

This study successfully developed and validated a deep learning model for the preoperative grading of lung cancer pathology. The model's accurate predictions could serve as a useful adjunct in treatment planning for lung cancer patients, enabling more effective and customized interventions to improve patient outcomes.