Deep learning synthesis of DBT features from mammography for breast cancer diagnosis

Digital breast tomosynthesis (DBT) enhances diagnostic accuracy by minimizing tissue overlap seen in digital mammography (DM). However, the substantial number of images generated by DBT poses challenges for radiologists. The aim of this study was to enhance the diagnostic accuracy and clinical utility of DM for breast cancer diagnosis by leveraging advanced feature representations derived from DBT.

This retrospective study was approved by the institutional review board of the institute, and samples were consecutively collected from February 2019 to August 2020. Our dataset comprised 806 samples (training: n = 570; testing: n = 236). We developed a cross-modality feature synthesis (CMFS) model to generate synthetic DBT (sDBT) features from DM images. To adeptly capture and analyze structures of varying sizes, such as tumors and their surrounding glandular regions, we designed a ResNet-based multiscale attention network (MSAN). Our generative adversarial network (GAN) model employs multitask learning to convert mammography features into sDBT features for breast cancer diagnosis.

Our DBT-based model achieved an area under the curve (AUC) of 0.916, utilizing a 32-channel attention bias and the MSAN architecture. Furthermore, the prediction model using the sDBT features generated from DM enhanced lesion classification, improving the AUC from 0.845 to 0.878 (p = 0.002) when relying solely on DM data. Ablation studies validated the crucial components of our model, underscoring its efficacy in breast cancer diagnosis. Model visualizations demonstrated precise lesion localization, and our method enhances the diagnostic value of DM by utilizing DBT data, potentially reducing patient radiation exposure.

The proposed CMFS framework improves breast cancer diagnosis by generating sDBT features from DM, enhancing diagnostic accuracy while reducing radiation exposure. Code and models are available at https://github.com/ibml2024/CMFS.