- Python 3+
- TensorFlow
- Keras
- Unrar
To download the dataset, download the files on my private google drive and copy them into dataset/raw3DVolumes/
To be able to download, you should have access to the drive and be connected on your account.
https://drive.google.com/uc?export=download&id=17QRHX5JSmqliQv_x6S1yUMhk_Sx5TqvI
copy the data-mvct.rar in the folder dataset/ and decompresses it into the folder raw3DVolumes (automatically created with the following code)
unrar e data-mvct.rar raw3DVolumes/
rm data-mvct.rar
You can now pre process these raw 3D volumes with the python script preProcessData.py that will generate the exploitable dataset.
In this project, our goal is to build an autoencoder in order to improve state-of-the-art image registration algorithm. This algorithm uses subsampled versions of images to register in order to compute a deformation flow, but we expect that by using more wise features we could improve the results. In this part of the work, we built a convolutional autoencoder that is able to retrieve the information contained in an image, by producing multiple image features of low resolution.
Below an outline of the architecture used :
To train this model, we first used a sample of our dataset composed of 6000 images. Using only 600 of them can allow us to get a good accuracy in a short time (~15min). Need access to the server to get the plot of the error on pretraining. Then, we train the model on the full dataset :
Here is an exemple of the reconstrution, with the fully trained model :
