Using convolutional neural networks to discriminate between cysts and masses in Monte Carlo-simulated dual-energy mammography.

Purpose: A substantial percentage of recalls (up to 20%) in screening mammography is attributed to extended round lesions. Benign fluid-filled breast cysts often appear similar to solid tumors in conventional mammograms. Spectral imaging (dual-energy or photon-counting mammography) has been shown to discriminate between cysts and solid masses with clinically acceptable accuracy. This work explores the feasibility of using convolutional neural networks (CNNs) for this task.

Methods: A series of Monte Carlo experiments was conducted with digital breast phantoms and embedded synthetic lesions to produce realistic dual-energy images of both lesion types. We considered such factors as nonuniform anthropomorphic background, size of the mass, breast compression thickness, and variability in lesion x-ray attenuation. These data then were used to train a deep neural network (ResNet-18) to learn the differences in x-ray attenuation of cysts and masses.

Results: Our simulation results showed that the CNN-based classifier could reliably discriminate between cystic and solid mass round lesions in dual-energy images with an area under the receiver operating characteristic curve (ROC AUC) of 0.98 or greater.

Conclusions: The proposed approach showed promising performance and ease of implementation, and could be applied to novel photon-counting detector-based spectral mammography systems.