Custom Dataset class for medical image patches

vision
1

Hi,

I’m trying to define a custom Dataset for my medical images. For each patient, I have a histopathology Whole Slide Images (WSI). I preprocess each image, obtaining a variable number of 224 x 224 patches and keep them in a folder with the patient ID. The label for each patch in a given patient folder is the label assigned to the WSI. The data directory structure is as follows:

image

I want to write a custom Dataset that will load each patch, with the correct label and feed it into DataLoader, in this case treating each patch as independent. My current implementation return a list of images and list of labels, as so:

class histoDataset(Dataset):

def __init__(self, main_dir, csv_file, transform):

    self.main_dir = main_dir 
    self.transform = transform 
    self.CSV = pd.read_csv(csv_file) 
    self.folder_id_code = self.CSV['Patient image ID'].tolist() 
    self.image_diagnosis = self.CSV['Classification.based.on.CD20'].tolist()
    self.classes = list(dict.fromkeys(self.image_diagnosis))
    self.label = self.CSV['Classification.based.on.CD20'].apply(self.classes.index)
    self.label_dict = dict(zip(self.label, self.classes))
    self.folders = [f.path for f in os.scandir(self.main_dir)]

def __len__(self):
     return len(self.folders)

def __getitem__(self, idx): 
    image_list = []
    image_labels = []
    self.folder_loc = os.path.join(self.main_dir, self.folder_id_code[idx])
    self.folder_files = [f.path for f in os.scandir(self.folder_loc)]
    for image in self.folder_files:
        image = Image.open(image) 
        image_tensor = self.transform(image) 
        image_label = self.label[idx] 
        image_list.append(image_tensor)
        image_labels.append(image_label)
        
    return image_list, image_labels

After inputting the resulting histoDataset into DataLoder and running this:

sample_batch = next(iter(df_loader))
sample_images, sample_labels = sample_batch
print_images = sample_images[0:10]
print_labels = sample_labels[0:10]

to check the images and labels, I get the following error:

Traceback (most recent call last):

  File "C:\Users\AmayaGS\Documents\PhD\Data\Dataset_loader.py", line 248, in <module>
    sample_batch = next(iter(df_loader))

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 521, in __next__
    data = self._next_data()

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 561, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\_utils\fetch.py", line 47, in fetch
    return self.collate_fn(data)

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\_utils\collate.py", line 84, in default_collate
    return [default_collate(samples) for samples in transposed]

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\_utils\collate.py", line 84, in <listcomp>
    return [default_collate(samples) for samples in transposed]

  File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\_utils\collate.py", line 82, in default_collate
    raise RuntimeError('each element in list of batch should be of equal size')

RuntimeError: each element in list of batch should be of equal size

Could someone please point out how to correctly implement the histoDataset and why I am getting this error? Thanks!

2

Maybe something is wrong in the transform of your dataset. Can you post it here?

1 Like
3

This is how I use my transform

import torchvision
import torchvision.transforms as transforms

transforms = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

df = histoDataset(main_dir, csv_file, transforms)

df_loader = torch.utils.data.DataLoader(df, batch_size=10, shuffle=True, num_workers=0, drop_last=False)

4

The transform looks good to me. The error message basically says that some of the loaded data are of different shapes. I think it should be image_list not image_labels.
Is it possible that self.folder_files can have different number of items when loading with different idx? For example, sometimes a folder contains 2 images, while sometimes it contains 4. Then the loaded data shape will be [2, 3, 224, 224] and [4, 2, 224, 224], which causes the problem.

5

I’m not sure if you structure this data loading correctly.
If I understand this right, the number of patches may be different from patient to patient.
Consequently, your image_list, image_labels variables that you return from getitem are of different length.
That might be the reason why you get: RuntimeError: each element in list of batch should be of equal size
As I don’t have this data and the code running in my notebook, this is just a vague hypothesis.

Did you consider getitem returning 1 patch only?