Slice-based versus Volumetric Deep Learning Classification for Breast Cancer Detection in Digital Breast Tomosynthesis

Digital Breast Tomosynthesis (DBT) is a form of three-dimensional imaging that reduces tissue overlap and improves lesion conspicuousness compared to conventional mammography. However, the vast number of slices per DBT study significantly increases the interpretation challenge and raises difficulties for computer-aided assessment. Herein, two deep learning methods of discerning breast cancer from DBT are compared: a 2D convolutional neural network (CNN), where each slice is studied in isolation, and a 3D CNN that exploits spatial information between slices directly. We do not wish through this comparison to create a benchmark between specific architectures, but instead examine the impact of volumetric feature learning on classification performance. Both models were trained and tested under the same circumstances with 2,000 DBT volumes from the publicly accessible BCS-DBT dataset, divided into 70% training, 15% validation, and 15% testing sets. The slice-based 2D CNN using a ResNet-18 backbone scored 86.7% accuracy, 85.3% sensitivity, 88.0% specificity, and 0.89 AUC under the receiver operating characteristic curve. By comparison, the volumetric 3D CNN performed better with 93.3% accuracy, 92.7% sensitivity, 94.0% specificity, and 0.95 AUC under the receiver operating characteristic curve. Analysis of the confusion matrix showed that the 3D CNN reduced both false negatives (22 to 11) and false positives (18 to 9) by half as compared to the 2D model. Notably, the improvements are gained at the expense of the 3D CNN being a shallower network compared to the 2D ResNet-18 network, underscoring that volumetric context plays a larger role in DBT analysis than network depth itself. These results prove that deep learning in the volumetric setting markedly enhances diagnostic ability in DBT. With fewer cancers missed and fewer unjustified recalls, the 3D CNNs present a more clinically stable basis for computer-aided diagnosis and hold a promising future for the workflow of screening for breast cancer.