The discrimination accuracy on the hold-out test for classifying subtypes on mammograms was 69.1% (47/68), when four ViT features, four global features, age, and orientation were used.  The discrimination accuracy for each of the four subtypes was 64.7% (11/17) for Luminal A, 76.5% (13/17) for Luminal B, 64.7% (11/17) for HER2-enriched, and 70.1% (12/17) for triple-negative (Fig. 5).  Figures 6 to 9 illustrate instances of accurately classified cases within each category. Consequently, the application of a model combining deep learning with ViT for feature extraction and GBM for classification allowed for the distinction of subtypes in mammograms.  Nevertheless, the performance of this model may not be fully realized due to limitations in the available training data.  We plan to continue further validation by increasing our database.  In addition, we used PCA followed by LightGBM.  However, further improvement can be expected by using a nonlinear feature selection model such as independent component analysis (ICA) or t-Distributed Stochastic Neighbor Embedding (t-SNE).

In this study, we utilized ViT features extracted from a deep learning model, global features derived from pixel values, and clinical features encompassing patient age, as well as left and right breast information. To evaluate their impact, we systematically toggled each feature on and off, and the optimal results were achieved with the current combination of features. The findings suggest that the ability to classify subtypes is enhanced through the complementary nature of the features obtained. Further enhancement in performance can be anticipated by incorporating additional information, such as genetic data, which has been discussed in correlation with breast cancer, and patient-related factors like BMI, among other details currently unavailable.

The classification of triple-negative and other subtypes holds paramount significance, particularly in identifying patients who warrant early intervention.  When the number of ViT and global features were changed to focus on the triple negative, 88.2% (15/17) were obtained when the ViT and global features had 4 and 64 dimensions, respectively (Fig. 10).  This outcome is comparable to the accuracy observed in vacuum-assisted biopsy (VAB) when utilized as a substitute for surgical specimens. Of particular note is the non-invasive nature of mammography; should these findings be validated in a large-scale database, they could carry significant clinical implications.