Abstract
Deep learning can bring time savings and
increased reproducibility to medical image analysis. However,
acquiring training data is challenging due to the
time-intensive nature of labeling and high inter-observer
variability in annotations. Rather than labeling images, in
this work we propose an alternative pipeline where images
are generated from existing high-quality annotations using
generative adversarial networks (GANs). Annotations
are derived automatically from previously built anatomical
models. Annotations are transformed into realistic
synthetic ultrasound images with paired labels using a
CycleGAN. The pipeline developed is fully extensible to
any 2D or 3D segmentation or landmark detection task in
any modality. We demonstrate the pipeline by generating
synthetic 2D echocardiography images to compare with
existing deep learning ultrasound segmentation datasets.
A convolutional neural network is trained to segment the
left ventricle and left atrium using only synthetic images.
Networks trained with synthetic images produce accurate
segmentations on real images with median Dice scores of
0.90, 0.89, and 0.88 for left ventricle segmentation of three
different unseen datasets. These results match or are better
than inter-observer results measured on real ultrasound
datasets and are comparable to a network trained on a
separate set of real images. The proposed pipeline opens
the door for automatic generation of training data for many
tasks in cardiac imaging. The source code and models are
available to other researchers
increased reproducibility to medical image analysis. However,
acquiring training data is challenging due to the
time-intensive nature of labeling and high inter-observer
variability in annotations. Rather than labeling images, in
this work we propose an alternative pipeline where images
are generated from existing high-quality annotations using
generative adversarial networks (GANs). Annotations
are derived automatically from previously built anatomical
models. Annotations are transformed into realistic
synthetic ultrasound images with paired labels using a
CycleGAN. The pipeline developed is fully extensible to
any 2D or 3D segmentation or landmark detection task in
any modality. We demonstrate the pipeline by generating
synthetic 2D echocardiography images to compare with
existing deep learning ultrasound segmentation datasets.
A convolutional neural network is trained to segment the
left ventricle and left atrium using only synthetic images.
Networks trained with synthetic images produce accurate
segmentations on real images with median Dice scores of
0.90, 0.89, and 0.88 for left ventricle segmentation of three
different unseen datasets. These results match or are better
than inter-observer results measured on real ultrasound
datasets and are comparable to a network trained on a
separate set of real images. The proposed pipeline opens
the door for automatic generation of training data for many
tasks in cardiac imaging. The source code and models are
available to other researchers
| Original language | English |
|---|---|
| Journal | Ieee Transactions on Medical Imaging |
| Publication status | Accepted/In press - 12 Jan 2021 |