Figure 5
From: Denoising diffusion probabilistic models for 3D medical image generation
Visualization of the breast segmentation performance for six different studies (rows). The first two columns show the original MR image and the ground truth segmentation of the breast. The third column shows the segmentation of the Swin UNETR neural network when only 5% of the available data from the internal dataset is used during training. The fourth column shows the segmentation of the Swin UNETR when pre-trained in a self-supervised approach using 2000 synthetic images generated using a dataset from another institution and finetuned based on only 5% of the available data from the internal dataset. Green areas denote correctly segmented areas. In contrast, red areas denote deviations from the ground truth.
