Preoperative CT-based Deep Learning Model for Predicting Disease-Free Survival in Patients with Lung Adenocarcinomas
Menée en Corée du Sud à partir de données portant sur 800 patients atteints d'un adénocarcinome du poumon de stade T1-2aN0M0 (âge médian : 64 ans) et sur 108 patients supplémentaires (âge médian : 63 ans), cette étude évalue la performance d'un modèle d'apprentissage automatique, utilisant des données issues de tomographies numériques pré-opératoires, pour prédire la survie sans maladie des patients
Background : Deep learning models have the potential for lung cancer prognostication, but model output as an independent prognostic factor must be validated with clinical risk factors.
Purpose : To develop and validate a preoperative CT-based deep learning model for predicting disease-free survival in patients with lung adenocarcinoma.
Materials and Methods : In this retrospective study, a deep learning model was trained to extract prognostic information from preoperative CT examinations. Data set 1 for training, tuning, and internal validation consisted of patients with T1–4N0M0 adenocarcinoma resected between 2009 and 2015. Data set 2 for external validation included patients with clinical T1–2aN0M0 (stage I) adenocarcinomas resected in 2014. Discrimination was assessed by using Harrell C index and benchmarked against the clinical T category. The Greenwood-Nam-D’Agostino test was used for model calibration. The multivariable-adjusted hazard ratios (HRs) were analyzed with clinical prognostic factors by using the Cox regression.
Results : Evaluated were 800 patients (median age, 64 years; interquartile range, 56–70 years; 450 women) in data set 1 and 108 patients (median age, 63 years; interquartile range, 57–71 years; 60 women) in data set 2. The C indexes were 0.74–0.80 in the internal validation and 0.71–0.78 in the external validation, both comparable with the clinical T category (0.78 in the internal validation and 0.74 in the external validation; all P > .05). The model exhibited good calibration in all data sets (P > .05). Multivariable Cox regression revealed that model outputs were independent prognostic factors (hazard ratio [HR] of the categorical output, 2.5 [95% confidence interval {CI}: 1.03, 5.9; P = .04] in the internal validation and 3.6 [95% CI: 1.6, 8.5; P = .003] in the external validation). Other than the deep learning model, only smoking status (HR, 3.4; 95% CI: 1.4, 8.5; P = .007) contributed further to prediction of disease-free survival for patients after resection of clinical stage I adenocarcinomas.
Conclusion : A deep learning model for chest CT predicted disease-free survival for patients undergoing an operation for clinical stage I lung adenocarcinoma.
Radiology , résumé, 2019