Patient
Since this is a retrospective analysis, Jinhua Central Hospital's Ethics Committee, which is connected to Zhejiang University, has waived the need for patient-informed consent. A retrospective collection of pre-treatment 18F-FDG PET/CT imaging data, along with clinical and pathological data, was conducted for breast cancer patients confirmed by pathology at Jinhua Hospital affiliated with Zhejiang University, covering the period from August 5, 2015, to July 25, 2024. Data were rigorously screened according to inclusion and exclusion criteria. The patient selection process is illustrated in Fig. 1 .
Inclusion criteria: 1) No treatment prior to PET/CT examination, including surgery, radiotherapy, chemotherapy, and immunotherapy; 2) Both the primary breast cancer lesion and ALNM lesions were confirmed by pathology; 3) Complete clinical and pathological information.
Exclusion criteria: 1) Individuals who were on anticancer medication prior to PET/CT scanning; 2) Inadequate clinical and pathological information; 3) Existence of more cancerous tumors; 4) Poor image quality of PET/CT, affecting interpretation of results.
A training cohort and a testing cohort were randomly selected from the dataset in a 6:4 ratio. Instances from the testing cohort were used for an impartial assessment of the prediction model's performance, while all instances from the training cohort were used to train the model.
PET/CT Scanning Technology
All enrolled patients underwent scanning using the Siemens PET/CT-Biograph mCT. Prior to the examination, patients were told to fast for 6–8 hours, and their height, weight, and blood glucose levels were measured. An intravenous injection of the 18F-FDG imaging agent was administered at a dose of 0.1–0.15 mCi/kg, ensuring that the patient's fasting blood glucose level was below 11.1 mmol/L prior to injection. The patient was then instructed to rest quietly for 1 hour before undergoing PET/CT scanning.
The following were the CT scan parameters: 120 kV tube voltage and a tube current automatic mAs technology, the slice thickness is 3 mm, pitch is 0.8, and rotation duration is 0.5 s/r. The PET scan parameters were as follows: 1.5 min/bed, with 6 to 8 bed positions, and the slice thickness of 3 mm. For post-processing, the scanned photos are uploaded to the imaging workstation.
Image Assessment
The images are evaluated by two radiologists at our center, each with over three years of diagnostic experience, who are kept unaware of the patients' clinical information and pathology results. The evaluation criteria are as follows: A higher uptake of the radiopharmaceutical in breast tissue compared to surrounding areas indicates the presence of breast cancer, while an increased uptake in lymph nodes compared to adjacent muscle tissue suggests lymph node metastasis. A semi-automated delineation of the region of interest (ROI) is performed based on a 40% SUVmax threshold, measuring PET metabolic parameters, including SUVmax, minimum standardized uptake value (SUVmin), average standardized uptake value (SUVavg), metabolic tumor volume (MTV), and total lesion glycolysis (TLG). Additionally, the maximum diameter of the lesion in the axial plane is measured. In cases with multiple lesions, only the largest lesion by volume is selected for measurement.
Radiomics
Data Collection: Export the raw DICOM data from the "Medex" platform. Image Segmentation: Using the 3D Slicer software (version 5.6.1), select the "Draw" tool to manually delineate the ROI layer by layer on the PET/CT fused images, taking care to avoid necrotic areas. Perform radiomics feature extraction on the segmented ROI using the "Radiomics" package for both PET and CT images. Figure 2 displays the radiomics analysis workflow used in this investigation. Three groups of radiomics features can be distinguished: (I) geometry, (II) intensity, and (III) texture. Tumors are characterized by geometric features in three dimensions. The first-order statistical distribution of voxel intensities inside the tumor is described by intensity characteristics. The higher-order spatial distribution of the intensity patterns is described by the texture features. Z-score normalization is used to address the issue of scale variation in manually extracted radiomics features.
Using a double-blind technique, two nuclear medicine doctors separately segmented each lesion. The intraclass correlation coefficient (ICC) was used to assess the stability of feature extraction; an ICC > 0.75 indicates strong consistency in feature extraction.
Clinical and Pathological Characteristics
Clinical information were recorded for all included patients, including age, body mass index (BMI), menstrual status, tumor T stage, carcinoembryonic antigen (CEA) levels, carbohydrate antigen 125 (CA125) levels, and carbohydrate antigen 153 (CA153) levels. Additionally, the following pathological and immunohistochemical data were recorded: estrogen receptor (ER) status, progesterone receptor (PR) status, human epidermal growth factor receptor 2 (HER-2) status, tumor cell proliferation marker 67 (Ki-67) levels, and molecular subtypes of breast cancer. ER positivity is defined as ≥ 10% of tumor cell nuclei staining positive; PR positivity is defined as ≥ 10% of tumor cell nuclei staining positive, with ≥ 20% indicating high expression, and lower percentages indicating low expression; HER-2 positivity is defined as an immunohistochemical score of ≥ 2 + or positive FISH amplification; Ki-67 expression index ≥ 14% indicates high expression, while lower percentages indicate low expression. According to the 2023 NCCN guidelines[12], breast cancer molecular subtyping is determined based on immunohistochemical results, specifically: luminal A, luminal B, HER-2 positive, and triple-negative.
Feature Selection and Model Construction
A Mann-Whitney U test was run on all radiomics characteristics, preserving only those with a p-value < 0.05. The correlation between characteristics with high redundancy was determined using Pearson's rank correlation coefficient, which kept features with a correlation coefficient of at least 0.9 between any two features. For additional feature selection, the Least Absolute Shrinkage and Selection Operator (LASSO) was utilized. To find the ideal λ, 10-fold cross-validation was utilized, keeping features with non-zero coefficients. Finally, the Rad score was calculated by summing the products of the LASSO-Logistic regression coefficients and their corresponding values. This study also compared the stability and reliability of seven machine learning algorithms in predicting ALNM, including Logistic Regression (LR), Naive Bayes, K-Nearest Neighbors (KNN), Decision Tree, Random Forest (RF), Extra Trees, and Extreme Gradient Boosting (XGBoost). The diagnostic performance of the optimal predictive model was evaluated using the Receiver Operating Characteristic (ROC) curve.
Building the clinical model follows a nearly identical procedure to that of building the radiomics model. First, statistical variables with a p-value < 0.05 were selected from the clinical baseline data, and a clinical model was constructed using the optimal machine learning algorithm, incorporating 5-fold cross-validation and a fixed testing cohort. Finally, a nomogram was constructed based on logistic regression analysis by integrating radiomics features and clinical characteristics, and the ROC curve and calibration curve were plotted to evaluate its performance. The clinical utility was evaluated using DCA.
Statistical Analysis
The Python statsmodels module (version 0.13.2) was utilized for statistical analysis, and a p-value < 0.05 was considered statistically significant. For continuous variables, the Mann-Whitney U test was utilized to examine intergroup differences; for categorical variables, the χ2 test was employed.