Ultimate-Storm
commented
1 year ago - reconstruct the images to 3d, using dcm2niix to have nifti files, https://github.com/rordenlab/dcm2niix
- crop the tumor to the segmentation, slice the tumor and get jpg or png images from every slice and then we use that as tiles
- "Ax Dyn Pre" series would be optimal for training
In short: dcm -> nifti -> slice to jpg as tiles
- [x] dcm -> nifti
- [x] slice to jpg as tiles
Converted data for "Ax Dyn Pre" series stored at 192.168.33.102 /mnt/sda1/swarm-learning/radiology-dataset/converted-niix/ All series converted stored at 192.168.33.102 /mnt/sda1/swarm-learning/radiology-dataset/converted-niix-all-series/
Although there are ways just to convert dcm files for jpg format, then we could treat them just as tiles and do feature extraction. But there is a huge loss in image accuracy. Also after we convert them to 3d plane(niix format for each patient) we might be able to use the annotation box, also it would be optimal for 3D-CNN networks.
Daniel suggested we start with this dataset: https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70226903 these are N=922 breast MRI exams. The patients all have cancer, but we have many labels available (survival, subtype, etc.). So one idea would be to use this dataset, split it into four partitions, and then use SL to train a classifier for categorical targets. However, downloading the data requires a special tool and will cost 30 mins of work. In the end, use the 3Dslicer software (https://www.slicer.org/) to look at the images.