Page History

Demo of SynthStrip

• Dr. Malte Hoffman's slides (request an accessible version)
• SynthStip prevents facial reconstruction from medical images with pixel-level de-identification.
• Combined effort led by Andrew at the Martinez Center.
• He will present the tool and how it works and then will focus on its evaluation and robustness.
• Skull stripping removes non-brain tissue from CT scan images. The reason we do this is because irrelevant information can appear in downstream analysis algorithms. If we remove these structures, we can increase security. It's a holistic approach to de-id and de-facing.
• Other tools make strong assumptions about the type of image you can use them on. They are dependent on the type of image, like MRI scans. Resolution expectations.
• Implementations of skull stripping from deep learning have limitations too because they train. They would not do as well on new contrast types or new modalities.
• The goal is to get these neural networks to generalize to data that is unseen in training. We suggest to synthesize images of arbitrary characteristics and get these images to train label maps to do the synthesis.
• The SynthStrip team's general process is to start from a set of brain label maps but no images. Given an input label map, they create a new label map by flying a random nonlinear formation to increase spatial variability. Then they sample a right-scale image with a different intensity distribution for every structure in the image, supplying random artifacts such as blurring.
• Since this process is completely randomized, if they were to input the same underlying segmentation map again, they would get a completely different image. They want to encourage their networks to generalize beyond a specific data type.
• For brain extraction, they sample and augment a label map and synthesize an arbitrary image from it. This image is fed into a neural network. Then they compute the ground truth and compare. 
• Label maps are only needed for training. If you want to extract the brain from the image, you don't need the label map at test time.
• At evaluation time, we find that SynthStrip performs with robustness and high accuracy on MRI scans, isotropic scans that have fixed license, and diffusing weighted imaging that typically have lower resolution.
• Performance in the presence of pathology? SynthStrip did well against data from TCIA.
• To quantify this, they compiled a database of 600 images and created two metrics. For both metrics, SynthStrip outperforms all of the other baselines they tested. It avoids cross-mislabeling.

Brian Bialecki on De-identification

• His team at ACR is trying to find a way to release this data publicly.
• He would like to get patient consent to share some identifying data.
• They'd like to see the real data to assess both the real world risk of doing nothing and the real world risk of various mitigation approaches. 

Other Discussion

• Skull stripping is not a replacement for for de-facing.
• The value of SynthStrip is not in how well the model performs, but rather how the model is created.
• It's difficult to find things that work across different data sets and modalities. This is something we desperately need.
• SynthStrip does not work on slices, it is fully 3D.
• Access model, registered or restricted, for data with a data use agreement. This is what everyone is converging on.
• Record and track who got the data. But some repositories have no such tracking.

Next Meeting

• Skipping July
• August 9, 2022 at 1 p.m. EST