Cross-Device Adaptation of Mirai for Mammography-Based Breast Cancer Risk Prediction

The research focused on the Mirai model, which uses deep learning to predict breast cancer risk from mammographic images. It was tested on a large patient cohort with images from Hologic and General Electric (GE) mammography systems, including GE Premium View and GE Tissue Equalization. The native Mirai model showed lower performance on GE Tissue Equalization images compared to Hologic or GE Premium View images. When researchers fine-tuned the model on GE Tissue Equalization images, its performance on these images improved, but performance on Hologic images substantially decreased. To address this problem, they developed a device-invariant model using interleaved multi-device sampling and conditional adversarial training. This approach restored performance on Hologic images while maintaining improvements on GE Tissue Equalization images. The results demonstrate the importance of balanced domain-adaptation strategies when deploying models across different clinical imaging environments.