Preoperative discrimination of invasive and non-invasive breast cancer using machine learning based on automated breast volume scanning (ABVS) radiomics and virtual touch quantification (VTQ)

Abstract

Purpose: Evaluating the efficacy of machine learning for preoperative differentiation between invasive and non-invasive breast cancer through integrated automated breast volume scanning (ABVS) radiomics and virtual touch quantification (VTQ) techniques.

Methods: We conducted an extensive retrospective analysis on a cohort of 171 breast cancer patients, differentiating them into 124 invasive and 47 non-invasive cases. The data was meticulously divided into a training set (n = 119) and a validation set (n = 52), maintaining a 70:30 ratio. Several machine learning models were developed and tested, including Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM). Their performance was evaluated using the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC), and visualized the feature contributions of the optimal model using Shapley Additive Explanations (SHAP).

Results: Through both univariate and multivariate logistic regression analyses, we identified key independent predictors in differentiating between invasive and non-invasive breast cancer types: coronal plane features, Shear Wave Velocity (SWV), and Radscore. The AUC scores for our machine learning models varied, ranging from 0.625 to 0.880, with the DT model demonstrating a notably high AUC of 0.874 in the validation set.

Conclusion: Our findings indicate that machine learning models, which integrate ABVS radiomics and VTQ, are significantly effective in preoperatively distinguishing between invasive and non-invasive breast cancer. Particularly, the DT model stood out in the validation set, establishing it as the primary model in our study. This highlights its potential utility in enhancing clinical decision-making processes.

Keywords: Automated breast volume scanning; Breast cancer; Machine learning; Radiomics; Virtual touch quantification.

Fig. 1

The flow chart for patient selection

Fig. 2

Workflow for the radiomics process

Fig. 3

Lasso regression curves. a The regression analysis was conducted with tenfold cross-validation to select the optimal radiomic features for differentiating invasive from non-invasive breast cancer. b Distribution graph of regression coefficients for the optimal radiomic features

Fig. 4

Distribution of radiomics score. a Training set; b Validation set

Fig. 5

ROC curves of each univariate and combined model. a Training set; b Validation set

Fig. 6

presents the feature importance analysis based on SHAP values for the DT model. Panel a Represents the Training Set, and Panel b Represents the Validation Set. The figure clearly indicates that Radscore is the most influential feature in this study's DT model, significantly driving the model's decision-making process. Other features, such as coronal plane features and SWV, also contribute to the model's predictions, but their impact is less pronounced compared to Radscore

Fig. 7

ROC Curves of the Imaging Model, Radiomics Label Model, and Combined Model with the DT Algorithm as the Output Model. Panel a Represents the Training Set, and Panel b Represents the Validation Set