Computed Tomography-based Prognostication in Lung Adenocarcinomas through Histopathological Feature Learning: A Retrospective Multicenter Study.

Abstract

Rationale: Modeling imaging surrogates for well-validated histopathological risk factors would enable prognostication inearly-stage lung adenocarcinomas. Objectives: We aimed to develop and validate computed tomography (CT)-based deep learning (DL) models for the prognostication of early-stage lung adenocarcinomas through learning histopathological features and to investigate themodels' reproducibility using retrospective, multicenter datasets. Methods: Two DL models were trained to predict visceral pleural invasion and lymphovascular invasion, respectively, using preoperative chest CT scans from 1,426 patients with stage I-IV lung adenocarcinomas. The averaged model output was defined as the composite score and evaluated for the prognostic discrimination and its added value to clinicopathological factors in temporal (n = 610) and external test sets (n = 681) of stage Ilung adenocarcinomas. The study outcomes were freedom from recurrence (FFR) and overall survival (OS). Interscan and interreader reproducibility were analyzed in 31 patients with lung cancer who underwent same-day repeated CT scans. Results: For the temporal test set, the time-dependent area under the receiver operating characteristic curve was 0.76 (95%confidence interval [CI], 0.71-0.81) for 5-year FFR and 0.67 (95% CI, 0.59-0.75) for 5-year OS. For the external test set, the area under the curve was 0.69 (95% CI, 0.63-0.75) for 5-year OS. The discrimination performance remained stable in 10-year follow-up for both outcomes. The prognostic value of the composite score was independent of and complementary to the clinical factors (adjusted per-percent hazard ratio for FFR [temporal test], 1.04 [95% CI, 1.03-1.05; P < 0.001]; OS [temporal test], 1.03 [95% CI, 1.02-1.04; P < 0.001]; OS [external test], 1.03 [95% CI, 1.02-1.04; P < 0.001]). The likelihood ratio tests indicated added value of the composite score (all P < 0.05). The interscan and interreader reproducibility were excellent (Pearson's correlation coefficient, 0.98 for both). Conclusions: The CT-based composite score obtained from DL of histopathological features predicted survival in early-stage lung adenocarcinomas with high reproducibility.