Reading png images and labels in Monai

[Harris91-lee]

Hello everyone, looks like LoadImaged is used for .nii.gz 3D volume: LoadImaged(keys = ["image", "label"]), how can i do that for only 2D png images?

[KumoLiu]

Hi @Harris91-lee, we have PILReader which can be used in LoadImage.

MONAI/monai/data/image_reader.py

Line 1129 in ff43028

class PILReader(ImageReader):

Hope it helps.

[KumoLiu]

reader="PILReader" should be ok.

MONAI/monai/transforms/io/array.py

Lines 138 to 144 in ff43028

	reader: reader to load image file and metadata
	- if `reader` is None, a default set of `SUPPORTED_READERS` will be used.
	- if `reader` is a string, it's treated as a class name or dotted path
	(such as ``"monai.data.ITKReader"``), the supported built-in reader classes are
	``"ITKReader"``, ``"NibabelReader"``, ``"NumpyReader"``, ``"PydicomReader"``.
	a reader instance will be constructed with the `*args` and `**kwargs` parameters.
	- if `reader` is a reader class/instance, it will be registered to this loader accordingly.

[Harris91-lee]

Thank you SO much for your fast reply @KumoLiu , you are a life saver.
so the correct form is: LoadImaged(keys=["image", "label"], image_only=False, reader="PILReader", reverse_indexing=False), and if i load only images no label: LoadImaged(keys="image", image_only=True, reader="PILReader", reverse_indexing=False)?

[KumoLiu]

Yes, feel free to ask if meet any issue. Thanks.

[Harris91-lee]

so there is something wrong in loading the images with their labels, the images tesnor shape is [1,1,512,512] which is as expected but the label comes like [1,512,512,512] though im using the same process for both:
class Custom_Dataset(CacheDataset):
def init(self, data_path, is_label = True, select_label = None, crop_size = None , input_size = None, augmentationMode = None, cache_rate = 0):
dt_data = open(data_path, 'r').readlines()
dt_data = [l.strip().split() for l in dt_data]
self.filenames_input = [os.path.join(data_dir, item[0]) for item in dt_data]
self.is_label = is_label
if is_label:
self.filenames_label = [os.path.join(data_dir, item[1]) for item in dt_data]
self.select_label = select_label # here it says true.

    self.crop_size = crop_size
    self.input_size = input_size

    data_files = self.__datafiles__()
    augmentation = self.__augmentation__(augmentationMode)
    super().__init__(data = data_files, transform = augmentation, cache_rate = cache_rate)

# 1. setup database
def __datafiles__(self):
    filenames_lst = []
    for file_name_dir in self.filenames_input:
        file_name = os.path.basename(file_name_dir)
        # file_name = file_name.replace('.nii.gz', '')
        file_name = file_name.replace('_image.png', '')

        filenames_lst.append(file_name)

    if self.is_label:
        data_files = [{"filename": filename, "image": img, "label": seg} for filename, img, seg in zip(filenames_lst, self.filenames_input, self.filenames_label)]
    else:
        data_files = [{"filename": filename, "image": img} for filename, img in zip(filenames_lst, self.filenames_input)]
    
    return data_files
    
    Do you see where is my mistake @KumoLiu ?  

[Harris91-lee]

Oh so i solved it by doing EnsureChannelFirstd(keys=["image", "label"]), instead EnsureChannelFirstd(keys="image"),

[Harris91-lee]

Hello Mr YunLiu @KumoLiu, hope you are doing well :)
i have a further question regarding the LoadImaged(keys=["image", "label"], image_only=False, reader="PILReader") how can i enforce reading images in gray so channel will be always 1? currently my data is mixed of gray and RGB so i have conflict regarding that.
Thank you so much in advance.

[Harris91-lee]

converting the images to gray externally is a solution i know but i wonder if i can do it internally ^_^

[KumoLiu]

Hi @Harris91-lee,
There isn't a direct transform function for converting RGB images to grayscale.
However, you can create a custom transform using common Python libraries such as OpenCV or PIL, and integrate it into your MONAI pipeline.
Here is a simple example of a custom MONAI transform:

from monai.transforms import Transform
from PIL import Image

class RgbToGray(Transform):

    def __call__(self, img):
        gray_img = Image.fromarray(img.astype('uint8')).convert('L')
        return np.array(gray_img)

The Transform base class is leveraged here to create a custom transform that can be integrated into a MONAI workflow. In this case, the transformation converts an RGB image to a grayscale image using the PIL library's Image.convert('L') method. This custom transform can then be incorporated into your MONAI pipeline like any other MONAI transform.
Please keep in mind that this transform assumes that data coming in are numpy arrays of image data. Depending on what your actual pipeline looks like, you might need to adjust the transform.

Hope it helps, thanks.

[Harris91-lee]

Thank you @KumoLiu.
that was indeed a great way, will use it ^^
During the training process i noticed my model size is so huge close to 1GB, is it normal to have this big size model? Are there any options I should add/remove if i want to reduce the model size?
I'm so grateful for your help and support though.

[Harris91-lee]

Hello @KumoLiu, hope you are doing well?
go tanother question regarding the LoadImaged(keys=["image"], image_only=True, reader="PILReader")
i noticed it does random/90 degree rotation to the data, is that normal? and how to stop that from happening?
many thankls in advance.

• Harris

[KumoLiu]

Hi @Harris91-lee, you can set reverse_indexing=False, this is enabled by default, so that output of the reader is consistent with the other readers.

MONAI/monai/data/image_reader.py

Lines 1136 to 1138 in e9e2738

	reverse_indexing: whether to swap axis 0 and 1 after loading the array, this is enabled by default,
	so that output of the reader is consistent with the other readers. Set this option to ``False`` to use
	the PIL backend's original spatial axes convention.

Thanks.

[Harris91-lee]

Thank you @KumoLiu :)
you are right, after adding the reverse_indexing=False no rotation happens but surprisingly the model result got worsen as the train dataloader didnt have reverse_indexing=False. Thought it's okay if the rotation happened in the training but seems i should unify the LoadImaged for both right?

[Harris91-lee]

So when i unify the train and val LoadImaged to have reverse_indexing=False the result comes out fine.
Thank you again @KumoLiu ~