Construction of a comprehensive predictive model for axillary lymph node metastasis in breast cancer: a retrospective study

Patients with suspected BC undergoing PET/CT at our hospital, from November 2016 to April 2022, were selected via picture archiving and communication as well as hospital information systems, based on the inclusion process depicted in Fig. 1. This study was conducted in accordance with the principles of the Declaration of Helsinki and was reviewed and approved by the Medical Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University(No. IRB-SOP-AF-16). All data were anonymized prior to analysis. Tumor staging was done in accordance with the eighth edition of the American Joint Committee on Cancer staging manual [11].This study was funded by the Department of Science and Technology of Shaanxi Province(No. 2023-YBSF-480), and registered with ClinicalTrials.gov (Date of first registration: 24/04/2023, ClinicalTrials.gov Identifier: NCT05826197).

Inclusion criteria: (1) ¹⁸ F-FDG PET/CT for breast lesions; (2) women with pathologically confirmed BC (age ≥ 18 years); (3) no history of surgery, radiotherapy, or chemotherapy before ¹⁸ F-FDG PET/CT; and (4) interval between ¹⁸ F-FDG PET/CT and puncture/surgery ≤ 2 weeks.

Exclusion criteria: (1) multifocal, bilateral, or occult BC; (2) incomplete clinical or pathological data; (3) poor PET/CT image quality preventing automated segmentation of metabolic tumor volume (MTV); and (4) concomitant malignant tumors.

PET/CT was performed on all patients using a 64-detector scanner (Gemini TF PET/CT, Philips, Netherlands). ¹⁸ F-FDG was synthesized by GE MINItrace mini cyclotron and Tracerlab FX-FDG synthesizer, and the synthetic precursor kit was purchased from ABX, Germany. The synthesized ¹⁸ F-FDG was released with a purity of ≥ 95%, and the quality was assured to be suitable for human injection. Patients fasted for at least 6 h before injection and had a fasting blood glucose level of less than 12.0 mmol/L. ¹⁸ F-FDG (dose 370 MBq/kg) was intravenously injected from the contralateral upper extremity of the affected mammary gland. The patients were encouraged to have sufficient water intake and rest for 60 min. The parameters for the CT scans were as follows: tube voltage 120 kV, tube current 300 mA, layer thickness 5 mm, layer spacing 5 mm, 512 × 512 matrix. PET collected 7–10 beds with 1.5 min/bed. PET images were corrected by the same machine used for CT data attenuation and reconstructed using an iterative method and time of flight. The imaging data were transferred to a workstation for image post-processing.

The PET/CT center’s chief physician and senior attending physician reviewed the images together and disagreement, if any, was resolved by consensus. The lesion was visually identified. A 3D region of interest (ROI) of the lesion was automatically outlined using the 40% threshold method, and PET metabolic parameters were measured, seen as Fig. 2.

Breast lesions with radionuclide concentrations greater than those in normal breast tissue were considered to be BC lesions, while lymph nodes with radionuclide concentrations greater than those in muscle tissue were considered to be metastatic lymph nodes.

Image segmentation was performed using ITK-SNAP software [12] (version 3.6.0, http://​www.​itksnap.​org/​); Brush Style: circular, Brush Size: 10, Brush Options: 3D. The entire tumor volume was outlined on the PET image as ROI for segmentation, seen as Fig. 3. The lesions were marked by the attending physician and checked by the chief physician.

An open source Python package (PyRadiomics version 3.0.1 [13]) was used to extract the radiomics features from the ROI, and a total of 851 radiomics features were finally computed. These features were extracted and defined in accordance with the Image Biomarker Standardization Initiative.

Breast imaging reporting and data system classification was used to classify all BCs involving lymph nodes. The histological grading of BC was assessed using the internationally accepted Nottingham tissue grading system [14]. BC specimens were fixed in 4% formaldehyde solution and embedded in paraffin wax, sectioned at a thickness of 4 μm. They were routinely stained with HE and then subjected to immunohistochemistry which included evaluation of estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2, p53, and cell proliferation nuclear antigen Ki67.

Statistical analysis was performed using R language (version 4.1.0. R Foundation for Statistical Computing, Vienna, Austria, URL https://​www.​R-project.​org/​) and SPSS® (version 25.0, IBM Corp, Armonk, NY, USA) software with a significance level of α = 0.05. The data obtained were expressed as mean ± standard deviation. Groups were compared using the independent samples t-test if they were normally distributed with equal variance, otherwise the Mann–Whitney U test was used for comparison. Categorical variables were compared using the χ² test or Fisher’s exact test. The least absolute shrinkage and selection operator (LASSO) was used to downscale the radiomics features, and the Radiomics Score (RS) was established based on the coefficients of the downscaled features. Univariate and multivariate binary logistic regressions were used to construct three parametric models based on clinicopathology and ultrasound (model 1); clinicopathology, ultrasound, and PET (model 2); and clinicopathology, ultrasound, PET, and radiomics (model 3), for predicting LNM in BC. ROC curves were plotted and area under the curve (AUC) was calculated to evaluate the discrimination of the three models, and the AUCs of the three models were compared using the Delong test. Furthermore, 1000 bootstraps with put-back repeated sampling were used to internally validate the differentiation of the models and to calculate the corrected AUC. Calibration curves were plotted separately to evaluate the calibration of the three models. The net reclassification index (NRI) and integrated discrimination improvement index (IDII) were used to evaluate the inter-model improvement. Finally, decision curves were plotted to evaluate the net benefit of the three models for all patients.

A total of 124 BC patients with a median age of 49 years-old (20–76 years) were included in this study, and the clinic-pathological characteristics of the axillary LNM negative group (n = 89) and the axillary LNM positive group (n = 35) were compared to identify potential diagnostic biomarkers of axillary LNM. The T stage, US_BIRADS, US_LNM, and PET_LNM in the positive axillary LNM group was significantly higher than that of in the negative LNM group (P = 0.013, P = 0.049, P < 0.001, P < 0.001, respectively), seen as Table 1. There were no statistical differences in age, tumor location, quadrant distribution, subtypes, grade, mol-subtypes, SUVmax, SUVmean, SD and MTV between the two groups (P>0.05), as shown in Table 1.

The radiomics features were normalized using Z-scoring and then downscaled using LASSO regression, and the optimal lnλ = -2.704 was determined by cross-validation, as shown in Fig. 4. Based on the linear weighting of the four radiomics features and their coefficients, radiomics score for predicting LNM(RS_LNM) was calculated, that was, Zoriginal_firstorder_10Percentile * 0.0891130 + Zoriginal_glszm_SizeZoneNonUniformityNormalized * 0.2768424 + ZwaveletLLH_firstorder_Skewness * 0.1603961 + ZwaveletHHH_glrlm_ ShortRunEmphasis * 0.1117953–0.9779143. RS_LNM for the group of negative LNM and the group of positive LNM were − 1.090 ± 0.448 and − 0.693 ± 0.344, respectively, with a statistically significant difference (t = -4.720, P < 0.001; Fig. 5). In the ROC analysis, the AUC was 0.767 (95% CI: 0.674–0.861; Fig. 6).

The results showed that T stage, US_LNM, and PET_LNM were associated with RS_LNM. Three models with multivariate were constructed for predicting LNM, as shown in Table 2. RS_LNM was an independent predictor when US_LNM and PET_LNM were integrated in the multivariable model, with OR value of 8.078 [95%CI, (1.862–35.050), P < 0.05]. The differentiation of the three models was shown in Table 3. The ROC curves showed that model 3 outperformed model 1 for the sensitivity (model 3 vs. model 1, 82.86% vs. 48.57%), and outperformed model 2 for the specificity (model 3 vs. model 2, 82.02% vs. 68.54%). The AUC of mode 1, model 2 and model 3 was 0.687, 0.826 and 0.874, seen as Fig. 7, and the Delong test showed the AUC of model 3 was significantly higher than that of model 1 and model 2, seen as Fig. 8. The nomogram was the visualization of the model 3, seen as Fig. 9.

The calibration curves for all three models showed good calibration, seen as Fig. 10. The continuous NRI for model 2 relative to model 1 was 1.118 (95% CI: 0.797–1.422), p < 0.001, and IDII of 0.141 (95% CI: 0.078–0.203), which mean positive improvement. The continuous NRI for model 3 relative to model 2 was 0.666 (95% CI: 0.368–0.963), p < 0.001, and IDII of 0.060 (95% CI: 0.007–0.114), p = 0.026, which mean improvement.

Decision curve analysis showed that model 2 resulted in a higher degree of net benefit for all the patients than model 1, and model 3 resulted in a higher degree of net benefit for all the patients than model 2, seen as Fig. 11.