Open blotero opened 9 months ago
Add MORPHO MNIST support https://github.com/dccastro/Morpho-MNIST
Add Longitudinal Multiple Sclerosis Lesion Segmentation Challenge support: https://smart-stats-tools.org/lesion-challenge-2015
Add BRATS support:
https://www.sciencedirect.com/science/article/pii/S0031320323001012?via%3Dihub
www.braintumorsegmentation.org
www.virtualskeleton.ch/
For BRATS 2021: https://www.synapse.org/#!Synapse:syn25829067/wiki/610863
For BRATS 2023: https://www.synapse.org/#!Synapse:syn51156910/files/
Add LIDC-IDRI (lung abnormalities) support:
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3041807/
https://wiki.nci.nih.gov/display/CIP/LIDC
https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=1966254
QSMSC at UCL (real-world multiple sclerosis lesion dataset) does not seem to be a publicly available dataset. I've contacted first author and now I find myself waiting for a response.
Add LIDC-IDRI (lung abnormalities) support:
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3041807/
https://wiki.nci.nih.gov/display/CIP/LIDC
https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=1966254
Apparently, a tool is required for downloading the whole 125GB of images: https://wiki.cancerimagingarchive.net/display/NBIA/Downloading+TCIA+Images
I think we could use the tf native MNIST dataset instead of Morpho. This way, we can apply a similar disturbance strategy to the one applied to oxford pet instead of applying morphological transformations. https://www.tensorflow.org/api_docs/python/tf/keras/datasets/mnist
Add support for tomato seeds: https://gcpds-image-segmentation.readthedocs.io/en/latest/notebooks/02-datasets.html#Tomato-Seeds
Add support for tomato seeds: https://gcpds-image-segmentation.readthedocs.io/en/latest/notebooks/02-datasets.html#Tomato-Seeds
This dataset is way too small for being meaningful, a typical UNet actchitecture training will have loss curves like the following:
This is unacceptable for achieving decent annotators emulation.
We will be using the augmented version with detection target instead. However, it would still require to adapt it as tf generator dataset.
Add support for tomato seeds: https://gcpds-image-segmentation.readthedocs.io/en/latest/notebooks/02-datasets.html#Tomato-Seeds
This dataset is way too small for being meaningful, a typical UNet actchitecture training will have loss curves like the following:
This is unacceptable for achieving decent annotators emulation.
We will be using the augmented version with detection target instead. However, it would still require to adapt it as tf generator dataset.
Successful results for tomato seeds disturbing model:
Implementation will be included in nexts PR's
Add API's for mapping more datasets