Study population

This study was approved by the institutional review board, and informed consent was waived given that this was a retrospective study. Clinical and imaging data of patients with solid NSCLC diagnosed by histopathology after radical surgical resection in our hospital from January 2014 to September 2019 were retrospectively reviewed. The inclusion criteria included the following: (1) patients with solid NSCLC who were surgically resected and confirmed by histopathology; (2) chest CT examination was performed within 1 month before surgery; (3) CT images could be downloaded from the picture archiving and communication system (PACS); (4) regular and complete follow-up records after surgery (at least 3 years); (5) the TNM pathological stage was stage I; and (6) the slice thickness of CT images was less than or equal to 1.5 mm. The exclusion criteria were as follows: (1) preoperative radiotherapy and chemotherapy; (2) patients with other malignant tumours; and (3) patients who were lost to follow-up.

A total of 459 patients with solid stage I NSCLC met the study requirements (male, 239; female, 220; mean age, 60.24 ± 10.20 years; range, 21–84 years). There were 72 cases in the progressive group and 387 cases in the nonprogressive group. The patients were randomly allocated to the training cohort and the internal validation cohort at a ratio of 7:3. A total of 321 patients were included in the training cohort, and 138 patients were included in the internal validation cohort. A total of 104 cases (male, 54; female, 50; mean age, 57.13 ± 9.91 years; range, 36–76 years) were collected from another hospital according to the same criteria as an external validation cohort, including 15 cases in the progressive group and 89 cases in the nonprogressive group. Figure 1

Fig. 1

Overall flow chart of the study. (A) The ROI extraction process, (B) DLS building process, (C) CM building process, (D) MDLR building process. ROI: region of interest; DLS: deep learning signature; CM: clinical model; MDLR: multimodal deep learning radiomics

Overall flow chart of the studyChest CT scan protocol and CT finding evaluation

Chest CT scans were conducted by one of the following scanners: Definition Force (Siemens, German), Siemens 16 (Siemens, German), Toshiba Aqilion (Toshiba, Japan), and GE Discovery 64 (GE, America). Spiral CT volume scanning technology was adopted, and the scanning parameters are detailed in Supplementary S1. Subjective CT findings evaluation was performed by two radiologists (with 10 and 15 years of experience in chest imaging diagnosis) independently evaluating the CT findings of lung cancer lesions as detailed in Supplementary S2.

Clinical information record and follow-up plan

Stage I included stages IA and IB according to the International Association for the Study of Lung Cancer (IASLC) 8th edition TNM stage [22]. The clinical and pathological information of patients was recorded, including smoking history, pathological types, operation method, pathological stage and serum tumour markers. The follow-up plan was as follows: (1) chest CT was reviewed every 6–12 months for the first 2 years after surgery and once a year thereafter; (2) if there were clinical symptoms, the corresponding site was examined; and (3) the end point of the study was the progression of the disease. Patients with no progression were followed up for 3 years or more. The definition of postoperative progress in 3 years of lung cancer was according to prior studies [23] and is detailed in Supplementary S3.

Data preprocessing

The data used in the pulmonary nodule CT image dataset were approved by the hospital ethics committee, and the privacy protection of patients met the requirements of the regulations. The dataset adopts the original data in digital imaging and communications in medicine (DICOM) format, image matrix 512 × 512, without any modification, editing or lossy compression. The images of each case were kept continuous and complete, without missing layers or split layers.

Venous phase images with a continuous cross-sectional plane were used. The DL model input CT image regions of interest (ROIs) data were constructed as follows: the ROIs were delineated by radiologists with ten years of experience in chest diagnosis. The process involved identifying the location and contour of the lesion, then constructing a rectangular box based on the lesion’s contour that encompasses the entire boundary of the lesion. Consequently, the ROI includes both the entire lesion information and the surrounding tumor information. Therefore, the selection of ROI is less influenced by the subjective experience of the clinicians and does not require the radiologists to precisely delineate the tumor boundaries. For detailed steps, please refer to Fig. 2.

Fig. 2

Chest CT image preprocessing. First, we selected all consecutive transverse slices of venous-phase CT images of solid pulmonary nodule lesions (a); second, a rectangular bounding box containing the whole region of the lesion was drawn in CT images by reader 1 using our in-house method developed based on MATLAB 2016 (b); third, this rectangular bounding box was applied to other layers of the lesion to crop all CT images (c); finally, these images were resized to 224 × 224 (d); among them, n represents the number of slices of a lesion

Building the DLS

To prevent overfitting, this study employed a transfer learning strategy during the training of the DL model. Pretrained ResNet 18 [24] network parameters on the ImageNet dataset were used as the initial model. Subsequently, a fine-tuning strategy was applied, where the gradients of the convolutional layers in the first two layers of ResNet 18 were frozen, and the model was adjusted using the preprocessed images of solid NSCLC, categorized into progression and non-progression groups from this study until the training epochs reached the specified parameters, at which point the model training was stopped. The DL experimental equipment was shown in Supplementary S4.

Due to the redundancy of features in DL, it can lead to overfitting of the model’s classification performance. Therefore, in this study, a ResNet18 network was used as a feature extraction network. The convolutional kernels in the network’s convolutional layers were used as feature extractors, with each kernel corresponding to a DL feature. Subsequently, feature extraction was performed on all images for each patient, and the DL features extracted from all images were averaged to obtain a set of DL features for a single patient. The ResNet18 network contained a total of 3, 904 convolutional kernels, resulting in 3, 904 DL features being extracted for each patient. The feature extraction process of DL is shown in Fig. 3. The detailed parameters of DL training is shown in Supplementary S5.

Fig. 3

Deep learning feature extraction. First, the preprocessed pictures were input into the trained deep learning model; second, the convolution kernel was used as the feature extractor to average the eigenvalues obtained after the picture passed through each convolution kernel; finally, 3, 904 deep learning features were extracted from each patient by stitching the eigenvalues extracted from each convolution kernel

To further reduce computational complexity and identify highly relevant features for the task, this study employed feature selection techniques such as the U test and the maximum relevanceon the initial set of 3, 904 deep features. Subsequently, the selected features were used to construct an extreme learning machine (ELM) based on integrated strategy classifier (Supplementary S6). This process ultimately allowed for the classification of early lung cancer progression and nonprogression, thereby creating a DL signature.

Building MDLR model

To perform a comprehensive analysis of the postoperative progression risk status in NSCLC, the following three steps were performed in the training cohort to build a MDLR model that combined DLS, clinical pathological characteristics, and subjective CT findings. First, we used Cohen’s kappa test to analyse the subjective CT findings between the two radiologists, with values of poor (0.00–0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), and near-perfect agreement (0.81–1.00). Second, the Wilcoxon rank-sum test, Fisher’s exact test, or Pearson’s Chi-square test was used to compare clinical variables between groups (gender and age), subjective CT finding and DLS. Finally, the factors with significant differences were selected as inputs for the ELM classifier to construct the MDLR model, enabling a robust analysis of the postoperative progression risk.

Statistical analysis

R3.0.1 (http://www.rproject.org) and Python 3.6 were used for all statistical analyses. The ROC analyses and decision curve analysis (DCA) were performed using “pROC,” and “dca.r,” respectively. A t test was used to compare age and longest diameter between the progressive and nonprogressive groups. Pearson’s Chi-square test was used to compare gender, emphysema, margin, lobulated sign, speculated sign, vacuole sign, air bronchogram sign, surgical type, pathological stage, smoking history, NSE, Cyfra21-1, CEA and CA199. Fisher’s exact test was used to compare the location and pathological type. Clinical and pathological factors and subjective CT findings were used to construct the clinical model through univariate analysis and ELM.

To assess the performance of each diagnostic model, the sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), and accuracy were determined using receiver operating characteristic (ROC) curve analysis. The DeLong test was used to compare the area under the curve (AUC) of the models. The MDLR of the validation and training datasets was evaluated by calibration. By quantifying the net benefit of patients at various threshold probabilities in the cohort, DCA was used to assess the clinical usefulness of the predictive models. P < 0.05 on both sides was regarded as statistically significant.