Instance Spine Segmentation

[oguzcanbekar]

Hi there firstly I would like to be appreciate for your great work I hope one day I can also help to improve your code~~

My first question is I am tryoing to use my spine datas to instance segmentation, I have some nii files and nrrd files which I was segment manually. With the convenience provided by Torchio, I am able to create my dataset easily but I have question at that point, my datas have generally [512,512,600] dimensions but I would like to upload them to model by [512,512] dimensions. I mean for each epoch I would like to train by just 1 slice image or with batch more slices. I hope I could explain myself clearly until you answer I will also search but hopefully we can figure it out my all problems until I finish my all code. For this purpose I opened this discussion and I would like to ask my questions from here. Thank you very much in advance.

Basically this is my first code too by Torchio:

def Dataset_load(base_path):

dataset_dir = Path(base_path)


images_dir = dataset_dir  / 'Img/'
labels_dir = dataset_dir  / 'Mask/'

image_paths = sorted(images_dir.glob('*.nii'))
label_paths = sorted(labels_dir.glob('*.nrrd'))
print(len(image_paths), len(label_paths))

assert len(image_paths) == len(label_paths)

CT = 'ct'
MASK = 'mask'
subjects = []
for (image_path, label_path) in zip(image_paths, label_paths):
    subject = tio.Subject(
        CT = tio.ScalarImage(image_path),
        MASK = tio.LabelMap(label_path),
    )

    subjects.append(subject)
dataset = tio.SubjectsDataset(subjects)
return dataset

dataset = Dataset_load("DATASET")

[fepegar]

Hi, @oguzcanbekar. I suggest you read carefully the documentation on Patch-based pipelines. You could use the Queue and patches of size (512, 512, 1) to train with a 2D model.