Could you please provide a training data format for custom dataset training?

[duducheng]

Hi Tang,

Thanks for your excellent work!

As title, could you please provide a training data format for custom dataset training? How can I preprocess my dataset to train my model?

Thanks!

Cheers, Jiancheng

[tanghaotommy]

Thanks for your interest!

The fastest way is to write your own dataset in PyTorch. If you take a look at dataset/brain_reader.py, you may follow its return types.

The brain_reader dataset returns an array of four elements. The first is the input 3D volume [1, depth, height, width] a float32 torch tensor. The second is a list of ground truth bounding box of objects in the volume [num_of_objects, 6], the six elements for each object is their z, y, x, depth, height, width. The third is a list of object category id, denoting the class for the corresponding bounding box in the second. The final one is a one-hot encoding segmention mask for each all classes, of shape [num_of_classes, depth, height, width].

[jmarsil]

Hi @tanghaotommy I also want to run your pipeline on an internal dataset at our institution. We currently have exported all DICOMS with binary masks to images in .nrrd format. What will I need to do, in addition to this, to train your model with our data?

Thanks for the help in advance!!