2.1 Study Population
This retrospective study was approved by the Institutional Review Board of the Southwestern Medical University Hospital (No. KY2022216), and the requirement for written informed consent was waived. All procedures were conducted in accordance with the 1975 Declaration of Helsinki and its later amendments[16].This study retrospectively analyzed patients who underwent pathologically con-firmed non-specific invasive BC at the Affiliated Hospital of Southwest Medical University (Luzhou, China) from February 2021 and February 2022. The entire patient records of this study were obtained from the Institutional Picture Archiving and Communication System (PACS, Carestream, Canada). All patients were desensitized to protect personal privacy and met the following inclusion and exclusion criteria.
Inclusion criteria: 1. pathologically diagnosed as non-specific invasive BC. 2.Received SLNB/ALND. 3. All patients performed unenhanced and biphasic (arterial and venous phase) contrast-enhanced CT before SLNB/ALND.
Exclusion criteria: 1. CT images were incomplete, or image quality was poor; 2. Patients were suffering from other tumors. 3. Incomplete clinicopathological data. 4. preoperative therapy (radiotherapy, chemotherapy, or other treatments).
A total of 780 patients with non-specific invasive BC were enrolled in the study, we excluded 584 patients, such as lack of complete three phase images (n = 539), clinicopathological incomplete data (n = 45). Finally, 196 patients met the criteria for inclusion and exclusion, including 97 with ALNM and 99 with non-ALNM. Patients were assessed for clinical node stage (cNx) by imaging and clinical examination. The primary outcome of this study was pathological node stage (pNx), which was determined based on SLNB or ALND. All patients were divided into training set (n = 136) and validation set (n = 60) in chronological order, patients who underwent CECT examination before November 31, 2021, were included in the training set, and the remaining patients were included in the validation set. The patient recruitment and workflow are shown in Fig. 1.
2.2 CECT Image Acquisition
All patients preoperatively received both unenhanced and biphasic (arterial and venous phase) contrast-enhanced chest CT examination (Netherlands, Philips Medical Systems). The scanning method was as follows: the contrast agent (Iohexol, 320mg/mL) was injected using a two-barrel high-pressure syringe through the median cubital vein of the patients (the dosage was 1.0 mL/kg, and the flow rate was 3.0 mL/s). The CT value of blood vessels at the level of the aortic arch was monitored after injection of contrast agent. Enhanced CT scans are automatically triggered when the CT value reaches around 250 HU.and venous phase scans were performed after a delay of 30 s. The scanning range extended from the level of the lower neck to the bottom of the thorax in the supine position.
2.3 Region of interest (ROI) Segmentation
The non-enhanced and biphasic (arterial and venous phase) contrast-enhanced images of study patients were obtained from Digital Imaging and Communications in Medicine (DICOM). For each CT phase, only one slice with the largest tumor area was chosen visually by the radiologist, The regions of interest (ROI) segmentation was performed from the original CT images using ITK-SNAP software (version 3.8.0, open source software; http://www.itksnap.org)[17]. All manual segmentation were performed by 2 practicing experienced radiologists (Y.W., with 11 years of CT imaging experience; Y.T., with 2 years of CECT imaging experience), followed by reader 1 (Y.W.) and reader 2 (Y.T.) repeated co-determination. They selected the mediastinal window (window width: 350, window level: 40) in ITK-SNAP to maintain approximately 1–2 mm from the tumor margin and delineate ROI along the tumor margin. Both radiologists were blinded to the patient's clinical data when they evaluated the multiphase CT images. Any discrepancies that arose were discussed to resolve the differences until they reached consensus. Then, 50 patients in the training set were randomly selected, and their ROIs were segmented again by reader 1 and reader 2 for the assessment of inter-reader reproducibility of radiomic features.
2.4 Features Extraction
The radiomics features and deep learning features were extracted from multiphase CT images of patients with BC, respectively. We used an open-source python-based package Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/) for image preprocessing and radiomics feature extraction of ROI images[18]. To make features reproducible, an interclass correlation coefficient (ICC) higher than 0.75 was considered credible.
Resnet 50 was used to extract deep learning features, The cropped single ROI area was copied into three images to fit the RGB three channels of the model, the input area was 224x224 pixels, and the transfer learning took the pretrained weights of Resnet 50 on the ImageNet dataset as the initial weights of the model, The model was fine-tuned using our data. Resnet50 was adjusted from the original multi-classification task to a binary classification task, we extracted the deep learning features from the last layer of resnet50, and the PAC algorithm further compressed the deep learning features.
We normalized all variables, correlations between features were measured using Spearman algorithm, If the correlation coefficient between the two features was greater than 0.9, one of them was excluded. The least absolute shrinkage and selection operator (LASSO) algorithm with 10-fold cross-validation was utilized to select the optimal features.
2.5 Development and Validation of DLRN
RBF-SVM was used to develop radiomics signature and deep learning signature. DLRN was also developed by adding Clinical independent predictor features on the basis of radiomics signature and deep learning signature. Receiver operating characteristic (ROC) curves were used to calculate the area under the curve (AUC) with 95% confidence interval (CI), sensitivity (SEN), specificity (SPE), accuracy (ACC), positive predictive value (PPV) and negative predictive value (NPV) to evaluate the performance of these models. We used the Hosmer-lemeshow goodness-of-fit test for DLRN and draw the calibration curve[15]. The Decision curve analysis (DCA) of all models was performed to quantify the net benefit of patients under different threshold probabilities in the cohort to assess the clinical value of the predictive models in our study [20].
2.6 Statistical Analysis
All statistical works in our study were conducted in R studio (version 4.1.1). Univariate analysis was used to select statistically significant clinical characteristics (p < 0.05). The independent t-tests or Wilcoxon rank sum tests was used to compare the continuous data, chi-square cross-tabulation test or Fisher’s exact test were used to compare the differences of categorical variables. The Delong test was used to compare the AUC between the models. Hosmer-lemeshow performs a goodness-of-fit test for DLRN, p > 0.05 indicates that model has excellent fitting effect.