Pathological selection and general information
A total of 851 cases of breast NWL in the background of dense breasts detected by conventional ultrasound were included in this study at Dongguan People's Hospital from May 30, 2018 to January 20, 2024. Figure 1 shows detailed information regarding the registration procedure. Inclusion criteria: (i) Diagnosed with breast NML by ultrasound. (ii) Dense breast tissue diagnosed by ultrasound or mammography. (iii) patients who have spontaneously undergo a core needle biopsy, vacuum-assisted biopsy, or surgery with complete clinical data and patients with confirmed pathological diagnosis. Exclusion criteria: (i) Lesions identified as mass-type breast lesions on ultrasound. (ii) Patients with non-density breast tissue as determined by ultrasound or mammography. (iii) A history of neoadjuvant chemotherapy, radiation therapy, or endocrine therapy for breast disease prior to surgery. (iv) the nature or pathological diagnosis of the unclear lesion. (v) Low image quality. This study was approved by the hospital ethics committee of Southern Medical University's No. 10th Associated Hospital (Tokyo People's Hospital) (G-2023027) and exempts informed consent from patients due to compliance with the principles of the Declaration of Helsinki.
A flow chart for patient registration.
Instruments and methods
A GE Logic E9 color Doppler ultrasound system with high frequency linear probe (L-16–5) is used with a mechanical index of 0.16. Ultrasound was performed by a radiologist with over 10 years of experience. The patient was placed in the supine position with the arm raised over the head at a 90° angle and fully exposed breast and both x-fossas. Patients were asked to head towards their side if necessary. For all cases identified as non-mass lesions (NMLs) in dense breast background via conventional ultrasound, 2D images were used to assess lesion size, uniformity of internal echoes, presence of trace imprints, posterior acoustic attenuation, and distortion of surrounding structures. Color Doppler imaging was used to assess the vascular properties of the lesions. Ultrasound images representing the longest axis of the lesion are selected, exported in DICOM format, and preserved the complete density in 16-bit grayscale, as configured in the Logic E9 system. To ensure reproducibility, all DICOM files retained the original acquisition parameters, including a dynamic range of 60 dB, time gain compensation (TGC) settings, and spatial composite configuration. Pixel values were linearly normalized before feature extraction [0, 1] Range using Min-Max scaling based on device-specific maximum quantization units. Pixel values were linearly normalized before feature extraction [0, 1] Ranges using Min-Max scaling based on device-specific maximum quantization units18,19.
Image data collection and processing
Ultrasound images were imported into ITK-SNAP using a single-blind method20 (v4.20, http://www.itksnap.org/) By two radiologists with over 10 years of experience. Tumor boundaries are drawn manually along the edges, with the region of interest (ROI) defined. This included spine, lobulation, and surrounding satellite or radial features. Using ROI margins as baseline, morphological adaptive segmentation was applied to enlarge the tumor area outside 1 mm, 2 mm, 3 mm, 4 mm, and 5 mm. This generated images of peri_1mm, peri_2mm, peri_3mm, peri_4mm, peri_5mm, and intratumoral regions (see Figure 2). Interclass correlation coefficients (ICC) were used to assess the consistency of feature extraction among observers. The ICC ratings were categorized as follows: <0.40 was considered "poor", 0.40-0.59 was considered "fair", 0.60-0.74-1.00.
NML image sketch. Ultrasound original image (a) Intratumor mask (b), and a 1mm abdominal tumor mask (c), 2 mm (d), 3 mm (e), 4 mm (f), and 5 mm (g).
Construction of radioactive models
The drawn ROI images have been imported into Piradiomicstwenty one (v3.0.1, https://github.com/aim-harvard/pyradiomics), open source software. Here, image function extraction was performed on ultrasound images using wavelet filters while retaining the original image. This includes texture features such as morphological features, primary functions, gray level size zone matrix (GLSZM), gray level co-occurrence matrix (GLCM), adjacent gray tone difference matrix (NGTDM), gray level dependent matrix (GLDM), and gray level execution length (GLRLM).
The extracted radioactive features were standardized using Z-score normalization. To alleviate the potential multicollinearity between radioactive features, t– Testing is performed using functions within the training set, p-value <0.05. Pearson correlation coefficients were calculated to assess interfunctional relationships, and features with correlation coefficients ≥0.9 were considered redundant and excluded. Additional sensitivity analysis was performed using Bonferroni correction (α = 0.05/851) to assess the need for multiple test corrections. Based on the least squares error (MSE) criteria, we combined minimum absolute contraction and selection operator (Lasso) regression with 10x cross-validation to select the optimal subset of radioactive features and further eliminated the Collinia variable by assigning zero-value coefficients. Each selected feature was multiplied by the corresponding regression coefficient and the resulting values were summed to generate a radiation score for each patient.
Building and Evaluating Predictive Models
Based on the optimal radiomic features selected, radioactive models of intratumoral and tumor tumors were constructed using random forests (RF), decision tree (DT), extra trees (ET), support vector machines (SVM), logistic regression (LR), stochastic gradient descent (SGD), K-Nearth Neighbors (KNN), Xgboost, Gradient Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Boosting Teer LightGBM, and Bayes. A 10x cross-validation was applied to identify the optimal parameters for each model, and receiver operating characteristics (ROC) curves were plotted to compare the area under the curve (AUC) of all 14 radiation models.
To ensure that the performance of the multivariate model is not primarily influenced by a single feature, univariate logistic regression is performed for each radioactive feature to assess its individual discriminating ability. p– Value as a metric. Features with univariate AUCs above 0.70 were considered strong independent predictors. Furthermore, we used Delong's test to reduce AUCs for optimal multivariate models (a) the highest performance univariate function (AUC>0.70) and (b) the multivariate model except for top-univariate functions. Statistically significant differences (p<0.05) indicates that a complete multivariate model provides excellent predictive performance compared to individual features.
Univariate and multivariate logistic regression analyses were performed using the clinical and ultrasound capabilities of the training group to identify independent predictors associated with NML, leading to the construction of clinical and ultrasound models. Radioactive models with the best predictive performance combined with independent clinical and ultrasound predictors were selected to construct the composite model. ROC curves were plotted and AUC, accuracy, sensitivity, and specificity of each model were calculated to assess the predicted effects of NML. Calibration curves and decision curve analysis (DCA) were employed to assess the goodness of fit and clinical value of the model. Clinical and ultrasound independent predictors, and optimal radioactive models, regression coefficients (Betaodds ratio (OR) for each variable was calculated. Multivariate logistic regression analysis was then performed to construct the nomogram.
Statistical methods
Statistical analyses were performed using R (version 4.2.0, https://www.r-project.org) and Python (version 3.10, https://www.python.org). Model structure and hyperparameter tuning were performed in Python using the “Scikit-Learn” (version 1.4.1) and “Xgboost” (version 2.0.3) libraries. Hyperparameter optimization and cross-validation were implemented via “GridsearchCV” to minimize overfitting and improve model accuracy. Ten-fold cross-validation was performed only within the training cohort for feature selection, hyperparameter tuning, and model optimization. Final performance metrics were assessed only in independent test cohorts that were not involved in any stage of model development.
Xgboost is an ensemble tree technique based on losses produced by weak decision tree learners. Xgboost was trained as a baseline model and then trained the final model with optimized hyperparameters. The hyperparameters in the XGBoost model were adjusted using Scikit-Learn's GridSearchCV function to perform 10x cross-validation. Hyperparameters selected for optimization include Learning_rate, Gamma, Max_depth, Subsamp, Min_Child_Weight, and N_ESTIMATORS. Scikit-Learn's GridsearchCV method with 10x cross-validation was also used to tune the hyperparameters for RF, DT, ET, SVM, LR, SGD, KNN, XGBOOST, ADABOOST, GBDT, CATBOOST, LightGBM, and Bayes.
Category data is displayed as a percentage (%) and compared using chi-squares (χ2) Test. Kolmogorov – Smirnov test was used to assess continuous data for health. Usually distributed data is displayed as mean ± standard deviation (\(x \pm s \)) and comparison using student use t-test. Pearson correlation analysis was used to select relevant radioactive functions. Receiver operating characteristics (ROC) curves were used to evaluate the predictive performance of each model and compared relative performance using Delong'test. Nomogram prediction models were constructed and optimal thresholds were determined by ROC curve analysis. Thresholds were defined by maximizing the Youden index. Based on the optimal cutoff values, area under the curve (AUC), accuracy, sensitivity, specificity, positive predictors, and negative predictors were calculated for each model. Calibration curves were employed to assess model consistency, and decision curve analysis (DCA) quantified net profits with varying threshold probability to assess the clinical significance of the model. Significance level was set at α=0.05 (both sided).
