There were 1692 patients who underwent pulmonary resection for lung cancer between January 2008 and December 2015. The following exclusion criteria were applied: lung cancer other than NSCLC, clinical stage IB–IV, incomplete surgical resection, wedge resection, no mediastinal lymph node dissection, and preoperative induction treatment. The remaining patients were those with clinical stage 0–IA NSCLC who underwent radical anatomical resection (lobectomy or segmentectomy) and systemic lymph node dissection at Tokyo Medical University Hospital. Among them, 720 patients for whom AI processing using their chest CT was successfully performed were enrolled in this study. We randomly assigned 480 and 240 patients to the derivation and validation sets, respectively. A consort diagram of patients included in the study was shown in Fig. 1. We reviewed the medical records of each patient for preoperative clinical information including TNM stage. The TNM stage was determined according to the eighth edition of the TNM classification of malignant tumors. The comorbidities included diabetes mellitus, cardiovascular disease, chronic obstructive pulmonary disease, cerebral disease, autoimmune disease, interstitial pneumonia, and asthma.
Patients were examined at the 6-month intervals for the first 2 years and 1-year intervals on an outpatient basis, with the aim of continuing follow-up for 10 years after resection. Follow-up evaluations included physical examinations, chest radiography, and blood tests. Chest and abdominal CT scans were performed every 6 months in the first 2 years and annually from the third year. Further evaluations, including brain magnetic resonance imaging and bone scintigraphy, were performed when the symptoms or signs of recurrence were observed. Positron emission tomography/computed tomography (PET/CT) was performed when appropriate. The date of recurrence was defined as the date of histologic proof or the date of identification based on clinicoradiologic findings by a physician.
Radiological evaluation of primary tumor
All patients in this study underwent preoperative high-resolution CT and three-dimensional CT lung modeling. Helical CT images (1.25-mm-thick) were obtained from the whole lung. The whole tumor and solid-part sizes were preoperatively measured by two experienced thoracic radiologists (Dr. R.M. with 14 years and Dr. J.P. with 30 years of experience in chest CT interpretation, respectively). The solid-part size was defined as the maximum dimension of the solid component of the lung window, excluding the ground-glass nodules.
Radiomics and AI imaging analysis of CT images
After the lung CT Digital Imaging and Communications in Medicine (DICOM) format data were transmitted to the Synapse Vincent system (Fujifilm Corporation, Tokyo, Japan), the AI software Beta Version (AI software; Fujifilm Corporation) in the system automatically detected and segmented lung nodules in the bilateral lungs and reconstructed the 3D images of the lungs and nodules. This segmentation algorithm was based on the 3D-Convolutional Neural Network using a modified U-Net architecture. The network consisted of 17 convolutional layers. The system separated the solid-part of lung nodules from non-solid-parts (ground-glass nodules; [GGNs]), and determined the size, volume, and ratio of solid-part, non-solid-part, and whole tumor lesions, as well as CT histogram features. The total of 39 AI imaging features included 22 radiomics features and 17 features from the GGN analysis, and the radiomics features were automatically extracted and displayed as a score from 0 to 1 using a feature analysis function. The 22 radiomics features were based on the labeling of 5118 tumors, and the datasets of the developmental process were divided into training, validation, and test sets. The trained model gave a mean area under the curve (AUC) score of 0.93 for all features on the test dataset (data not shown). This AI lung nodule analysis model uses a convolutional neural network based on VGG-16 and consists of 12 layers of convolution, with four layers removed from the output side of the VGG-16. To extract 3D radiomics features, 3D convolution was used for all the convolution layers.
Overall survival (OS) was measured from the day of surgery to the day of death from any cause or the day on which the patient was last known to be alive. Recurrence-free survival (RFS) was measured as the interval between the date of surgery and date of recurrence, date of death from any cause, or date the patient was last known to be alive. OS and RFS curves were plotted using the Kaplan–Meier method, and differences in variables were determined using the log-rank test. Univariate and multivariate logistic regression analyses were performed to identify the factors associated with pN using a Cox proportional hazards model. A backward stepwise selection method was used to build logistic regression models, and variables with a threshold of p < .15 were adopted for the stepwise model selection procedure. We conducted univariate and multivariate analyses separately using the 39 AI imaging features and other clinical factors. Pearson’s chi-square test (for categorical data) and Student’s t-test (for continuous data) were used to compare two groups of data. Receiver operating characteristics (ROC) curves for lymph node metastasis and early recurrence were constructed, and the optimal cut-off values were determined using the AUC. All tests were two-sided, and statistical significance was set at p < .05. The SPSS statistical software package (version 28.0, DDR3 RDIMM; SPSS Inc., Chicago, IL, USA) was used for statistical analysis. Violin plots were constructed using the R package (version 4.0.5).
The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. All procedures performed in this study involving human participants were performed in accordance with the Declaration of Helsinki (as reserved in 2013). The study was approved by the institutional review board of Tokyo Medical University (SH3951). Informed consent for the use and analysis of clinical data was obtained preoperatively for each patient.