A pre-processing tool to increase performance of deep learning-based CAD in digital breast Tomosynthesis

Purpose: Worldwide, female breast cancer is the fifth leading cause of death. Digital Breast Tomosynthesis (DBT) is increasingly involved in the routine diagnosis of breast cancer, providing quasi-three-dimensional reconstruction of the breast. DBT image analysis is time-consuming and Computed Aided Diagnosis (CAD) systems are becoming increasingly popular to automate DBT image analysis. Literature reports the importance of pre-processing operations on the final performance of the CAD itself. For this purpose, within the DeepLook project, we developed a pre-processing tool, called Digital Breast Imaging Tool (DBIT). Methods: DBIT procedure improves image contrast, removes breast skin contour and pectoral muscle, so that optimized datasets can be prepared to feed deep learning-based CAD. More than 200 DBT volumes were extracted from a public repository, equally divided into negative and positive tumors (both benign and malignant), to assemble the “raw dataset” (original images) and the “processed dataset” (images processed with DBIT). The classification performance of common convolutional neural networks (i.e. VGG16, ResNet18 and DarkNet19) on both raw and processed dataset was evaluated by the following metrics: AUROC, Accuracy, F1 score, Precision, Sensitivity and Specificity. Results: All neural networks performed better when trained and tested on the processed dataset, evidenced by the percentage increases in the mean values of all metrics considered. In particular, VGG16, ResNet18 and DarkNet19 presented an average classification accuracy increased by about 12%, 26% and 16%, respectively. Conclusions: Different deep learning-based CAD proved a significant reliability increase in breast cancer prediction, when trained and tested on images pre-processed with the proposed DBIT.