Development and Validation of Deep Learning Model Based on CT Simulator Images to Predict 5-Year Recurrence of Stage IIIA-N2 Non–Small-Cell Lung Cancer Patients Treated with Surgery and Postoperative Radiotherapy

Abstract

Purpose/Objective(s)Surgery followed by postoperative radiotherapy is a treatment choice for stage IIIA-N2 non–small cell lung cancer (NSCLC) patients recommended by NCCN guidelines. However, a large group of these patients develop recurrence within five years. Growing evidence suggests deep learning models constructed with computed tomography (CT) images have predictive power. Therefore, we developed and validated a deep learning model to predict stage IIIA-N2 NSCLC patients' recurrence within five years, to instruct treatment choices and follow-up strategies.Materials/MethodsThe CT simulator images from stage IIIA-N2 NSCLC patients, treated with surgery followed by postoperative radiotherapy between 2003-2015, were collected. A prognostic model for 5-year recurrence was developed, using a deep neural network (ResNet) to perform visual recognition and data analysis. Data containing different imaging scanners and protocols were used to create a robust model for the variations. The model's performance was expressed as accuracy and the area under the receiver operating characteristic curve (AUC), and assessed using cross-validation.ResultsA total of 212 patients were eligible. Further data curation resulted in a final cohort of 70 patients: the training dataset had 54 patients, and the validation dataset had 16 patients. The 5-year recurrence rate was 52.86%, and event rates were balanced between training and validation groups. Our model learned to classify patients with recurrence vs without recurrence within five years, with 92% accuracy and AUC = 0.81 in the validation group. The deep learning model performed better than other machine learning models.ConclusionThe deep learning model performed very well and accurately predicted the 5-year recurrence of stage IIIA-N2 NSCLC patients treated with surgery followed by postoperative radiotherapy. The model could support clinicians in treatment decision-making and follow-up planning. Surgery followed by postoperative radiotherapy is a treatment choice for stage IIIA-N2 non–small cell lung cancer (NSCLC) patients recommended by NCCN guidelines. However, a large group of these patients develop recurrence within five years. Growing evidence suggests deep learning models constructed with computed tomography (CT) images have predictive power. Therefore, we developed and validated a deep learning model to predict stage IIIA-N2 NSCLC patients' recurrence within five years, to instruct treatment choices and follow-up strategies. The CT simulator images from stage IIIA-N2 NSCLC patients, treated with surgery followed by postoperative radiotherapy between 2003-2015, were collected. A prognostic model for 5-year recurrence was developed, using a deep neural network (ResNet) to perform visual recognition and data analysis. Data containing different imaging scanners and protocols were used to create a robust model for the variations. The model's performance was expressed as accuracy and the area under the receiver operating characteristic curve (AUC), and assessed using cross-validation. A total of 212 patients were eligible. Further data curation resulted in a final cohort of 70 patients: the training dataset had 54 patients, and the validation dataset had 16 patients. The 5-year recurrence rate was 52.86%, and event rates were balanced between training and validation groups. Our model learned to classify patients with recurrence vs without recurrence within five years, with 92% accuracy and AUC = 0.81 in the validation group. The deep learning model performed better than other machine learning models. The deep learning model performed very well and accurately predicted the 5-year recurrence of stage IIIA-N2 NSCLC patients treated with surgery followed by postoperative radiotherapy. The model could support clinicians in treatment decision-making and follow-up planning.