search

MIC-DKFZ / medicaldetectiontoolkit

The Medical Detection Toolkit contains 2D + 3D implementations of prevalent object detectors such as Mask R-CNN, Retina Net, Retina U-Net, as well as a training and inference framework focused on dealing with medical images.
Apache License 2.0
1.29k stars 296 forks source link

Questions on how to organize training data and generate bounding box ground truth #55

Open zwb0 opened 5 years ago

zwb0 commented 5 years ago

Sorry for asking this question, My training set is CT images consisting of kidney and kidney tumor, whose ground truth is labeled in voxel level. Now I want to do 3D object detection of tumor. I'm wondering how should I organize my training data to feed into the network. And should I generate bounding box ground truth for tumors manually or your library supports to generate bounding box ground truth automatically? Thanks you in advance! Best wishes

CodingHHW commented 5 years ago

in the experiments has demo you can learn

delton137 commented 4 years ago

I'm trying to do the same thing... will report back if I figure out how..

bsmarine commented 4 years ago

I do not see a demo in experiments showing expected training data hierarchy input, specifically for bboxes and/or voxel-level ground truth (if possible). Anyone have luck figuring this out?

pfjaegerfb commented 4 years ago

you can run, experiments/toy_exp/generate_toys.py . It will create the toy data.

bsmarine commented 4 years ago

Found it, thank you!

MancaZerovnikMekuc commented 4 years ago

I am looking at the 3D LIDC example. Is it possible to specify the format of the data which is returned by preprocessing.py? I would like to avoid the preprocessing of the big LIDC dataset just to see how to prepare my data.

I saw this: "Images are saved as numpy arrays and the meta-information for the corresponding patient is stored as a line in the dataframe saved as info_df.pickle."

Just a small example and description of the data preparation for the 3D data would be really helpful. I looked and generated the toy example dataset, but I am interested in 3D data, patch loader and voxel-wise annotation.

@delton137 did you figure it out?