From 070fc09af3f4b71654015e5a347e578363b46d43 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Anna=20B=C3=B8gevang=20Ekner?= <s193396@dtu.dk>
Date: Tue, 11 Feb 2025 09:54:37 +0100
Subject: [PATCH] added support for 3D in data preparation functions
---
qim3d/ml/_data.py | 160 +++++++++++++++++++++++++++++++++++-----------
1 file changed, 123 insertions(+), 37 deletions(-)
diff --git a/qim3d/ml/_data.py b/qim3d/ml/_data.py
index 253da4a1..7bba1467 100644
--- a/qim3d/ml/_data.py
+++ b/qim3d/ml/_data.py
@@ -4,6 +4,7 @@ from PIL import Image
from qim3d.utils import log
import torch
import numpy as np
+import nibabel as nib
from typing import Optional, Callable
import torch.nn as nn
from ._augmentations import Augmentation
@@ -54,7 +55,7 @@ class Dataset(torch.utils.data.Dataset):
self.sample_targets = [file for file in sorted((path / "labels").iterdir())]
assert len(self.sample_images) == len(self.sample_targets)
- # checking the characteristics of the dataset
+ # Checking the characteristics of the dataset
self.check_shape_consistency(self.sample_images)
def __len__(self):
@@ -64,10 +65,22 @@ class Dataset(torch.utils.data.Dataset):
image_path = self.sample_images[idx]
target_path = self.sample_targets[idx]
- image = Image.open(str(image_path))
- image = np.array(image)
- target = Image.open(str(target_path))
- target = np.array(target)
+ if image_path.suffix in ['.nii', '.nii.gz']:
+
+ # Load 3D volume
+ image_data = nib.load(str(image_path))
+ target_data = nib.load(str(target_path))
+
+ # Get data from volume
+ image = np.asarray(image_data.dataobj, dtype=image_data.get_data_dtype())
+ target = np.asarray(target_data.dataobj, dtype=target_data.get_data_dtype())
+
+ else:
+ # Load 2D image
+ image = Image.open(str(image_path))
+ image = np.array(image)
+ target = Image.open(str(target_path))
+ target = np.array(target)
if self.transform:
transformed = self.transform(image=image, mask=target)
@@ -78,13 +91,13 @@ class Dataset(torch.utils.data.Dataset):
# TODO: working with images of different sizes
- def check_shape_consistency(self,sample_images: tuple[str]):
+ def check_shape_consistency(self, sample_images: tuple[str]):
image_shapes= []
for image_path in sample_images:
image_shape = self._get_shape(image_path)
image_shapes.append(image_shape)
- # check if all images have the same size.
+ # Check if all images have the same size
consistency_check = all(i == image_shapes[0] for i in image_shapes)
if not consistency_check:
@@ -97,41 +110,102 @@ class Dataset(torch.utils.data.Dataset):
)
return consistency_check
- def _get_shape(self,image_path):
- return Image.open(str(image_path)).size
+ def _get_shape(self, image_path):
+ full_suffix = ''.join(image_path.suffixes)
+
+ if full_suffix in ['.nii', '.nii.gz']:
+ # Load 3D volume
+ image = nib.load(str(image_path)).get_fdata()
+ return image.shape
+
+ else:
+ # Load 2D image
+ image = Image.open(str(image_path))
+ return image.size
-def check_resize(im_height: int, im_width: int, resize: str, n_channels: int) -> tuple[int, int]:
+def check_resize(
+ orig_shape: tuple,
+ resize: tuple,
+ n_channels: int,
+ is_3d: bool
+ ) -> tuple:
"""
- Checks the compatibility of the image shape with the depth of the model.
- If the image height and width cannot be divided by 2 `n_channels` times, then the image size is inappropriate.
- If so, the image is changed to the closest appropriate dimension, and the user is notified with a warning.
+ Checks and adjusts the resize dimensions based on the original shape and the number of channels.
Args:
- im_height (int) : Height of the original image from the dataset.
- im_width (int) : Width of the original image from the dataset.
- resize (str) : Type of resize to be used on the image if the size doesn't fit the model.
+ orig_shape (tuple): Original shape of the image.
+ resize (tuple): Desired resize dimensions.
n_channels (int): Number of channels in the model.
+ is_3d (bool): Whether the data is 3D or not.
+
+ Returns:
+ tuple: Final resize dimensions.
Raises:
ValueError: If the image size is smaller than minimum required for the model's depth.
"""
- # finding suitable size to upsize with padding
- if resize == 'padding':
- h_adjust, w_adjust = (im_height // 2**n_channels+1) * 2**n_channels , (im_width // 2**n_channels+1) * 2**n_channels
+
+ # 3D images
+ if is_3d:
+ orig_d, orig_h, orig_w = orig_shape
+ final_d = resize[0] if resize[0] else orig_d
+ final_h = resize[1] if resize[1] else orig_h
+ final_w = resize[2] if resize[2] else orig_w
+
+ # Finding suitable size to upsize with padding
+ if resize == 'padding':
+ final_d = (orig_d // 2**n_channels + 1) * 2**n_channels
+ final_h = (orig_h // 2**n_channels + 1) * 2**n_channels
+ final_w = (orig_w // 2**n_channels + 1) * 2**n_channels
- # finding suitable size to downsize with crop / resize
- else:
- h_adjust, w_adjust = (im_height // 2**n_channels) * 2**n_channels , (im_width // 2**n_channels) * 2**n_channels
-
- if h_adjust == 0 or w_adjust == 0:
- raise ValueError("The size of the image is too small compared to the depth of the UNet. Choose a different 'resize' and/or a smaller model.")
-
- elif h_adjust != im_height or w_adjust != im_width:
- log.warning(f"The image size doesn't match the Unet model's depth. The image is changed with '{resize}', from {im_height, im_width} to {h_adjust, w_adjust}.")
+ # Finding suitable size to downsize with crop / resize
+ else:
+ final_d = (orig_d // 2**n_channels) * 2**n_channels
+ final_h = (orig_h // 2**n_channels) * 2**n_channels
+ final_w = (orig_w // 2**n_channels) * 2**n_channels
+
+ # Check if the image size is too small compared to the model's depth
+ if final_d == 0 or final_h == 0 or final_w == 0:
+ msg = "The size of the image is too small compared to the depth of the UNet. \
+ Choose a different 'resize' and/or a smaller model."
+
+ raise ValueError(msg)
+
+ if final_d != orig_d or final_h != orig_h or final_w != orig_w:
+ log.warning(f"The image size doesn't match the Unet model's depth. \
+ The image is changed with '{resize}', from {orig_h, orig_w} to {final_h, final_w}.")
+
+ return final_d, final_h, final_w
- return h_adjust, w_adjust
+ # 2D images
+ else:
+ orig_h, orig_w = orig_shape
+ final_h = resize[0] if resize[0] else orig_h
+ final_w = resize[1] if resize[1] else orig_w
+
+ # Finding suitable size to upsize with padding
+ if resize == 'padding':
+ final_h = (orig_h // 2**n_channels + 1) * 2**n_channels
+ final_w = (orig_w // 2**n_channels + 1) * 2**n_channels
+
+ # Finding suitable size to downsize with crop / resize
+ else:
+ final_h = (orig_h // 2**n_channels) * 2**n_channels
+ final_w = (orig_w // 2**n_channels) * 2**n_channels
+
+ # Check if the image size is too small compared to the model's depth
+ if final_h == 0 or final_w == 0:
+ msg = "The size of the image is too small compared to the depth of the UNet. \
+ Choose a different 'resize' and/or a smaller model."
+
+ raise ValueError(msg)
+
+ if final_h != orig_h or final_w != orig_w:
+ log.warning(f"The image size doesn't match the Unet model's depth. \
+ The image is changed with '{resize}', from {orig_h, orig_w} to {final_h, final_w}.")
+ return final_h, final_w
def prepare_datasets(path: str, val_fraction: float, model: nn.Module, augmentation: Augmentation) -> tuple[torch.utils.data.Subset, torch.utils.data.Subset, torch.utils.data.Subset]:
"""
@@ -153,17 +227,29 @@ def prepare_datasets(path: str, val_fraction: float, model: nn.Module, augmentat
resize = augmentation.resize
n_channels = len(model.channels)
- # taking the size of the 1st image in the dataset
+ # Determine if the dataset is 2D or 3D by checking the first image
im_path = Path(path) / 'train'
first_img = sorted((im_path / "images").iterdir())[0]
- image = Image.open(str(first_img))
- orig_h, orig_w = image.size[:2]
-
- final_h, final_w = check_resize(orig_h, orig_w, resize, n_channels)
+ full_suffix = ''.join(first_img.suffixes)
+
+ # TODO: Support more formats for 3D images
+ if full_suffix in ['.nii', '.nii.gz']:
+ # Load 3D volume
+ image = nib.load(str(first_img)).get_fdata()
+ orig_shape = image.shape
+ is_3d = True
+
+ else:
+ # Load 2D image
+ image = Image.open(str(first_img))
+ orig_shape = image.size[:2]
+ is_3d = False
+
+ final_shape = check_resize(orig_shape, resize, n_channels, is_3d)
- train_set = Dataset(root_path = path, transform = augmentation.augment(final_h, final_w, augmentation.transform_train))
- val_set = Dataset(root_path = path, transform = augmentation.augment(final_h, final_w, augmentation.transform_validation))
- test_set = Dataset(root_path = path, split='test', transform = augmentation.augment(final_h, final_w, augmentation.transform_test))
+ train_set = Dataset(root_path=path, transform=augmentation.augment(final_shape, augmentation.transform_train))
+ val_set = Dataset(root_path=path, transform=augmentation.augment(final_shape, augmentation.transform_validation))
+ test_set = Dataset(root_path=path, split='test', transform=augmentation.augment(final_shape, augmentation.transform_test))
split_idx = int(np.floor(val_fraction * len(train_set)))
indices = torch.randperm(len(train_set))
--
GitLab