This retrospective single-center study was approved by our local institutional review board and written informed consent was waived.
2.1 Patient characteristics
We reviewed the contrast spectral CT data of 270 patients with breast disease admitted to our hospital from 2020 to 2022. All patients met the following inclusion criteria: (1) the first diagnosis of invasive breast cancer with a diameter of 5 cm or smaller; (2) no signs of ALN involvement palpable by physical examination; (3) treatment with SLNB or ALNs removal during the operation. Exclusion criteria were: (1) neoadjuvant systemic therapy (NST) before surgery and spectral CT; (2) pure ductal carcinoma in situ; (3) without treatment in our hospital; (4) poor image quality of the axillary region. Thus, based on these criteria, a total of 146 consecutive patients were included in this study. According to the postoperative pathological results, these patients were divided into a LANB group and a high axillary nodal burden (HANB) group (one or more macro-metastatic lymph nodes).
2.2 CT protocol
Non-enhanced and double-phase contrast enhanced CT examinations were performed on a 256 multi-detector row CT scanner (Revolution CT; GE Healthcare, Waukesha, Wis, USA) in dual-energy mode. All patients were scanned craniocaudally in the supine position with the bilateral arms elevated in close contact with the head. The whole chest was scanned from the superior aperture of the thorax to the inferior edge of the costophrenic angle, which covered the breast and axillary area. Nonenhanced scan and enhancement scans were conducted in the single source dual-energy CT mode by rapid switching of tube voltages between 80 kVp and 140 kVp. Double-phasic enhanced scans were performed after the patient was injected nonionic iodinated contrast agent (Iopamidol 400mg I/ml; Shanghai Bracco Sine Pharmaceutical, China) with a total dose of 50-60 ml by antecubital venous access at a rate of 2-3 mL/s followed by 30 mL saline flushing through a double-syringe power injector. The enhancement scans automatically began after the attenuation value of the descending aorta level reached the triggering threshold (up to 100 HU) in the arterial phase (AP), with 30 seconds delay in the venous phase (VP). The other scan parameters of the pre-defined protocol were as follows: tube current of 280 mA, collimation thickness of 1.25 mm, helical pitch of 0.992, and rotation speed of 0.28 seconds. The CT images were reconstructed automatically with an adaptive statistical iterative reconstruction algorithm (ASIR) by using GSI viewer software (GE Healthcare). All CT raw data were reconstructed into contiguous axial images with a section thickness of 1.25 mm. Two types of images were derived from the reconstruction of imaging for each patient: water- and iodine-based material decomposition images and a set of virtual monochromatic images at energies ranging from 40 to 140 keV.
2.3 Morphologic analysis
The morphologic of the ALN was performed on 70-keV monochromatic arterial and venous images (1.25 mm section width) [13] by using viewer software on a workstation (AW 4.7; GE Healthcare) with a multiplanar reconstruction technique. Two senior radiologists (with at least 10 years of breast imaging diagnostic experience) who were blinded to the pathologic results of the lymph nodes and breast analyzed the morphologic of ALNs. The ALNs were scored according to the most abnormal morphology, maximum cortical thickness, and hilum situation. The scoring criteria includes the following: shortest diameter (< 5mm, 0 point; ≥ 5mm, 1 point), a ratio of the longest to the shortest diameter (>2, 0 point; ≤2, 1 point), fatty hilum (evenly present, 0 point; unevenly present, 1 point; absent, 2 points), and nodal cortex thickness (<0.3 cm, 0 point; ≥0.3 cm, 1 point). The highest scored LN of patients were labeled. Any disagreements were resolved by consensus. Meanwhile, the largest diameter of the tumor was measured at 40keV-images because 40-keV monochromatic images have the higher iodine enhancement (IE) and contrast-to-noise ratio (CNR)values [14]. The location of tumors including superior inner quadrant (SIQ), lower inner quadrant (LIQ), superior lateral quadrant (SLQ), lower lateral quadrant (LLQ), central, and others were also recorded.
2.4 Quantitative Spectral CT Parameter Measurement
The GSI data were analyzed using the AW 4.7 workstation. The same two senior radiologists performed the quantitative spectral CT parameters by placing a region of interest (ROI) encompassing the entire labeled LN without fatty hilum and surrounding structures, combined with axial, sagittal, and coronal images. A round ROI was also placed on the descending aorta as a reference. The ROI placed on the 70-keV image was then propagated automatically by the GSI Viewer software to monochromatic images and to material decomposition images. The spectral CT parameters, including CT values measured on monochromatic images, the iodine concentration (IC) on iodine-based material decomposition images, the water concentration on water-based material decomposition images, and the effective atomic number (Zeff) value were calculated automatically. The normalized iodine concentration (nIC) and normalized Zeff (nZeff) were calculated using the following formula: nIC = IC the labeled LN / IC aorta and nZeff = Zeff-the labeled LN / Zeff-aorta. The slope of the spectral Hounsfield unit curve (λHU, in Hounsfield unit per kiloelectron-volt), which is defined as the difference between the CT value at 40 keV and that at 70 keV divided by the energy difference (30 keV), was calculated as follows: λHU = (HU40keV- HU70keV)/30 keV, where HU40keV represents the CT value measured on 40-keV images and HU70keV stands for the CT value measured on 70-keV images. To ensure consistency between the measurements, the dual-phase contrast-enhanced images were simultaneously loaded into the workstation, and the size, shape, and position of the ROIs in the two phases were kept the same by use of the copy and paste functions.
2.5 Pathologic Analysis
We reviewed and recorded the histopathological diagnostic data from the hospital’s electronic medical records. Each node was recorded as benign (stage pN0), isolated tumor cells (stage pN0[i+]; size ≤0.2 mm), micro-metastasis (stage pN1mi; size > 0.2 mm, and < 2.0 mm), or macro-metastasis (stage pN1–3; size > 2.0 mm). Estrogen receptor (ER) status, progesterone receptor (PR) status, HER2 status, Ki-67 index, and lymph node status were obtained from the surgical pathologic reports. Clinical data, including patient age menopausal status, and family history, were obtained from electronic medical records.
2.6 Statistical Analyses
All statistical analysis, model generation, and model evaluation were performed using statistical software SPSS (version 23.0) and R platform (version 4.0.3). The categorical variables using the c2 test or Fisher’s exact test. The distribution uniformity of the continuous parameters was evaluated by the Kolmogorov-Smirnov test. The mean ± standard deviation was used to represent the normally distributed data, and the median and interquartile range (IQR) was used to express non-normally distributed data. Continuous variables were compared between LANB and HANB groups by using Mann–Whitney U test (or Student’s t-test). A two-sided p-value <0.05 was considered statistically significant. Least absolute shrinkage and selection operator (LASSO) regression algorithm was used for feature selection after univariate analysis. The LASSO-logistic regression algorithm with penalty parameter tuning conducted by stratified 5-fold cross-validation was applied by shrinking the coefficients of useless features to zero with the regulation parameter λ. The optimal features subset with the least cross-validation binominal deviance was used for the logistic regression equation construction. Receiver operating characteristic curve (ROC) analysis was used to evaluate the prediction model in the training cohort and the validation cohort, respectively. Including calculation of the AUC value and 95% confidence interval (CI). Diagnostic performance metrics, including sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV) were reported in the training and testing cohorts. Delong test was used to measure the differences of ROC curves among the three prediction models. In order to access the clinical usefulness of the model, we quantified the net benefit at different threshold probabilities in the data set with decision curve analysis (DCA).