Skip to main content
Fig. 3 | European Radiology Experimental

Fig. 3

From: Diffusion probabilistic versus generative adversarial models to reduce contrast agent dose in breast MRI

Fig. 3

The reconstruction of a mass enhancing lesion (a) and a non-mass enhancing lesion (b) was performed by GAN and DDPM using 25%, 10%, and 5% reduced-dose images. Patient A was a 62-year-old woman with a triple-negative invasive breast cancer (blue arrow) in the left breast with a small satellite lesion ventral the main tumor (green arrow): no special type, Grade 3, stage pT1c (tumor size > 10 and ≤ 20 mm in diameter). Although the main lesion could be reconstructed even at 5% dose, details of its margin and the surrounding spicula are no longer discernible. The satellite lesion located ventral to the main tumor is still residual in the GAN 5% reconstruction, but no longer discernible in the DDPM 5% reconstruction. Vessels are no longer discernible in detail in any of the 10% and 5% reconstructions. Patient B was a 39-year-old woman with extensive segmental intermediate-grade ductal carcinoma in situ − DCIS (blue arrow) in the left breast. When comparing GAN 5% and DDPM 5%, the segmental enhancement appears marginally fainter in the DDPM 5% latter reconstruction. DDPM Denoising diffusion probabilistic model, GAN Generative adversarial network

Back to article page