Stacked-Capsule-Autoencoder (SCAE) is An unsupervised method that combines Transformers and Capsule Networks to exploit spatial relations of different parts of an image.
- Stacked Capsule Autoencoders - NeurIPS 2019
In this repository, we implement, train and test the suggested SCAE for bone age assessment task - RSNA Pediatric Bone Age Challenge (2017).
The original BoneAge DS is heavy - images are around 1MB so we used https://github.com/karayanni/data-processing for preprocessing to make SCAE reasonable (too heavy otherwise).