Detecting breast implants on mammograms: a small network, just 1.4 per cent the size of ResNet18, performs almost as well

At Emory University, four models were tested on 1,000 images. MammoCLIP achieved an AUROC of 0.999, but the tiny ResNetLite – which is the fastest – performed no worse than ResNet18.

Automatically identifying mammograms showing breast implants is used both to sort through large image databases for AI training and to adapt treatment to clinical routine. A team from Emory University, including Judy Wawira Gichoya and Hari Trivedi, compared two foundation models (RAD-DINO and MammoCLIP, whose representations are classified by an SVM) and two convolutional networks trained from scratch, ResNet18 and a lightweight version called ResNetLite.

All efficient, very different costs

The models were trained on 5,000 single-sided mammograms from the Emory Breast Imaging Dataset and tested on 1,000 hand-checked images. All of them achieved excellent results. MammoCLIP had the best AUROC (0.999) and the shortest training was 493 seconds. RAD-DINO detected implants best (sensitivity of 0.980) but was the slowest in both training and inference. ResNetLite, with 1.4% of the parameters of ResNet18, did not differ statistically significantly (AUROC of 0.993) and delivered the fastest predictions.

The authors conclude that the choice must depend on the context: a foundation model like MammoCLIP in a GPU-equipped hospital, a lightweight network for constrained or embedded deployments.