π₯πNew paper! Check out how we use diffusion models (same type of model used in #dalle2 and #Imagen ) to perform the detection of anomalies in the brain!
arxiv.org/abs/2206.03461
1/n
arxiv.org/abs/2206.03461
1/n
with @marksgraham_ Robert Gray @pfdacosta @PTudosiu @paulwright74 @jthteo @profgeraintrees Parashkev Nachev Sebastien Ourselin @mjorgecardoso and colleagues. In a collaboration between @KingsImaging and UCL, using the @nvidia Cambridge-1 supercomputer
2/n
2/n
In tinyurl.com/5eu59826, we found that using transformers for anomaly detection has a few limitations, e.g. limited context and exposure bias. Besides that, the segmentation of anomalies with transformers can take a long time to be performed, limiting its application.
3/n
3/n
Based on latent diffusion models (arxiv.org/abs/2112.10752 developed by @robrombach and @andi_blatt ), we create models that locate image regions with a high probability of being anomalous. Then, we inpaint these regions using the model knowledge about healthy data.
4/n
4/n
Using the difference between the inputted image and the inpainted one, we obtain competitive performance in synthetic anomalies as well in real MRI and CT images. With diffusion models and recent non-Markovian samplers, we speed up our method to just a few seconds!
5/n
5/n
Looking forward to sharing more details about our models on 3D anomaly detection and generative modelling experiments soon!
n/n
n/n
β’ β’ β’
Missing some Tweet in this thread? You can try to
force a refresh
γ
