def test_patient_metadata() -> None:
"""
Loading a dataset where all patient metadata columns are present
:return:
"""
file = full_ml_test_data_path("dataset_with_full_header.csv")
df = pd.read_csv(file, dtype=str)
subject = "511"
expected_institution = "85aaee5f-f5f3-4eae-b6cd-26b0070156d8"
expected_series = "22ef9c5e149650f9cb241d1aa622ad1731b91d1a1df770c05541228b47845ae4"
expected_tags = "FOO;BAR"
metadata = PatientMetadata.from_dataframe(df, subject)
assert metadata is not None
assert metadata.patient_id == subject
assert metadata.institution == expected_institution
assert metadata.series == expected_series
assert metadata.tags_str == expected_tags
# Now modify the dataset such that there is no single value for tags. Tags should no longer be
# populated, but the other fields should be.
df['tags'] = ["something", ""]
metadata = PatientMetadata.from_dataframe(df, subject)
assert metadata.series == expected_series
assert metadata.institution == expected_institution
assert metadata.tags_str is None
def test_sample_construct_copy(random_image_crop: Any, random_mask_crop: Any,
random_label_crop: Any) -> None:
sample = Sample(image=random_image_crop,
mask=random_mask_crop,
labels=random_label_crop,
metadata=PatientMetadata(patient_id='1'))
sample_clone = sample.clone_with_overrides()
assert sample.get_dict() == sample_clone.get_dict()
assert type(sample) == type(sample_clone)
sample_clone = sample.clone_with_overrides(metadata=PatientMetadata(
patient_id='2'))
assert sample_clone.patient_id == 2
def load_train_and_test_data_channels(
patient_ids: List[int],
normalization_fn: PhotometricNormalization) -> List[Sample]:
if np.any(np.asarray(patient_ids) <= 0):
raise ValueError("data_items must be >= 0")
file_name = lambda k, y: full_ml_test_data_path("train_and_test_data"
) / f"id{k}_{y}.nii.gz"
get_sample = lambda z: io_util.load_images_from_dataset_source(
dataset_source=PatientDatasetSource(
metadata=PatientMetadata(patient_id=z),
image_channels=[file_name(z, c) for c in TEST_CHANNEL_IDS],
mask_channel=file_name(z, TEST_MASK_ID),
ground_truth_channels=[file_name(z, TEST_GT_ID)]))
samples = []
for x in patient_ids:
sample = get_sample(x)
sample = Sample(image=normalization_fn.transform(
sample.image, sample.mask),
mask=sample.mask,
labels=sample.labels,
metadata=sample.metadata)
samples.append(sample)
return samples
def test_sample(random_image_crop: Any, random_mask_crop: Any, random_label_crop: Any, random_patient_id: Any) -> None:
"""
Tests that after creating and extracting a sample we obtain the same result
:return:
"""
metadata = PatientMetadata(patient_id=42, institution="foo")
sample = Sample(image=random_image_crop,
mask=random_mask_crop,
labels=random_label_crop,
metadata=metadata)
patched_sample = CroppedSample(image=random_image_crop,
mask=random_mask_crop,
labels=random_label_crop,
mask_center_crop=random_mask_crop,
labels_center_crop=random_label_crop,
metadata=metadata,
center_indices=np.zeros((1, 3)))
extracted_sample = sample.get_dict()
extracted_patched_sample = patched_sample.get_dict()
sample_and_patched_sample_equal: Callable[[str, Any], bool] \
= lambda k, x: bool(
np.array_equal(extracted_sample[k], extracted_patched_sample[k]) and np.array_equal(extracted_patched_sample[k],
x))
assert sample_and_patched_sample_equal("image", random_image_crop)
assert sample_and_patched_sample_equal("mask", random_mask_crop)
assert sample_and_patched_sample_equal("labels", random_label_crop)
assert np.array_equal(extracted_patched_sample["mask_center_crop"], random_mask_crop)
assert np.array_equal(extracted_patched_sample["labels_center_crop"], random_label_crop)
assert extracted_sample["metadata"] == extracted_patched_sample["metadata"] == metadata
def test_min_patient_metadata() -> None:
"""
Loading a dataset where only required columns are present
"""
df = pd.read_csv(full_ml_test_data_path("dataset.csv"), dtype=str)
df = df.drop(columns="institutionId")
patient_id = "1"
metadata = PatientMetadata.from_dataframe(df, patient_id)
assert metadata.patient_id == patient_id
assert metadata.series is None
assert metadata.institution is None
assert metadata.tags_str is None
def test_custom_collate() -> None:
"""
Tests the custom collate function that collates metadata into lists.
"""
metadata = PatientMetadata(patient_id='42')
foo = "foo"
d1 = {foo: 1, SAMPLE_METADATA_FIELD: "something"}
d2 = {foo: 2, SAMPLE_METADATA_FIELD: metadata}
result = collate_with_metadata([d1, d2])
assert foo in result
assert SAMPLE_METADATA_FIELD in result
assert isinstance(result[SAMPLE_METADATA_FIELD], list)
assert result[SAMPLE_METADATA_FIELD] == ["something", metadata]
assert isinstance(result[foo], torch.Tensor)
assert result[foo].tolist() == [1, 2]
def test_visualize_patch_sampling_2d(
test_output_dirs: TestOutputDirectories) -> None:
"""
Tests if patch sampling works for 2D images.
:param test_output_dirs:
"""
set_random_seed(0)
shape = (1, 20, 30)
foreground_classes = ["fg"]
class_weights = equally_weighted_classes(foreground_classes)
config = SegmentationModelBase(should_validate=False,
crop_size=(1, 5, 10),
class_weights=class_weights)
image = np.random.rand(1, *shape).astype(np.float32) * 1000
mask = np.ones(shape)
labels = np.zeros((len(class_weights), ) + shape)
labels[1, 0, 8:12, 5:25] = 1
labels[0] = 1 - labels[1]
output_folder = Path(test_output_dirs.root_dir)
image_header = None
sample = Sample(image=image,
mask=mask,
labels=labels,
metadata=PatientMetadata(patient_id='123',
image_header=image_header))
heatmap = visualize_random_crops(sample,
config,
output_folder=output_folder)
expected_folder = full_ml_test_data_path("patch_sampling")
expected_heatmap = expected_folder / "sampling_2d.npy"
# To update the stored results, uncomment this line:
# np.save(str(expected_heatmap), heatmap)
assert np.allclose(heatmap, np.load(
str(expected_heatmap))), "Patch sampling created a different heatmap."
assert len(list(output_folder.rglob("*.nii.gz"))) == 0
assert len(list(output_folder.rglob("*.png"))) == 1
actual_file = output_folder / "123_sampled_patches.png"
assert_file_exists(actual_file)
expected = expected_folder / "sampling_2d.png"
# To update the stored results, uncomment this line:
# expected.write_bytes(actual_file.read_bytes())
if not is_running_on_azure():
# When running on the Azure build agents, it appears that the bounding box of the images
# is slightly different than on local runs, even with equal dpi settings.
# It says: Image sizes don't match: actual (685, 469), expected (618, 424)
# Not able to figure out how to make the run results consistent, hence disable in cloud runs.
assert_binary_files_match(actual_file, expected)
def test_get_all_metadata(default_config: ModelConfigBase) -> None:
df = default_config.get_dataset_splits().train
assert PatientMetadata.from_dataframe(df, '1') == PatientMetadata(
patient_id='1', institution="1")
assert PatientMetadata.from_dataframe(df, '2') == PatientMetadata(
patient_id='2', institution="2")
def test_visualize_patch_sampling(test_output_dirs: TestOutputDirectories,
labels_to_boundary: bool) -> None:
"""
Tests if patch sampling and producing diagnostic images works as expected.
:param test_output_dirs:
:param labels_to_boundary: If true, the ground truth labels are placed close to the image boundary, so that
crops have to be adjusted inwards. If false, ground truth labels are all far from the image boundaries.
"""
set_random_seed(0)
shape = (10, 30, 30)
foreground_classes = ["fg"]
class_weights = equally_weighted_classes(foreground_classes)
config = SegmentationModelBase(should_validate=False,
crop_size=(2, 10, 10),
class_weights=class_weights)
image = np.random.rand(1, *shape).astype(np.float32) * 1000
mask = np.ones(shape)
labels = np.zeros((len(class_weights), ) + shape)
if labels_to_boundary:
# Generate foreground labels in such a way that a patch centered around a foreground pixel would
# reach outside of the image.
labels[1, 4:8, 3:27, 3:27] = 1
else:
labels[1, 4:8, 15:18, 15:18] = 1
labels[0] = 1 - labels[1]
output_folder = Path(test_output_dirs.root_dir)
image_header = get_unit_image_header()
sample = Sample(image=image,
mask=mask,
labels=labels,
metadata=PatientMetadata(patient_id='123',
image_header=image_header))
expected_folder = full_ml_test_data_path("patch_sampling")
heatmap = visualize_random_crops(sample,
config,
output_folder=output_folder)
expected_heatmap = expected_folder / ("sampled_to_boundary.npy"
if labels_to_boundary else
"sampled_center.npy")
# To update the stored results, uncomment this line:
# np.save(str(expected_heatmap), heatmap)
assert np.allclose(heatmap, np.load(
str(expected_heatmap))), "Patch sampling created a different heatmap."
f1 = output_folder / "123_ct.nii.gz"
assert_file_exists(f1)
f2 = output_folder / "123_sampled_patches.nii.gz"
assert_file_exists(f2)
thumbnails = [
"123_sampled_patches_dim0.png",
"123_sampled_patches_dim1.png",
"123_sampled_patches_dim2.png",
]
for f in thumbnails:
assert_file_exists(output_folder / f)
expected = expected_folder / ("sampled_to_boundary.nii.gz"
if labels_to_boundary else
"sampled_center.nii.gz")
# To update test results:
# shutil.copy(str(f2), str(expected))
expected_image = io_util.load_nifti_image(expected)
actual_image = io_util.load_nifti_image(f2)
np.allclose(expected_image.image, actual_image.image)
if labels_to_boundary:
for f in thumbnails:
# Uncomment this line to update test results
# (expected_folder / f).write_bytes((output_folder / f).read_bytes())
if not is_running_on_azure():
# When running on the Azure build agents, it appears that the bounding box of the images
# is slightly different than on local runs, even with equal dpi settings.
# Not able to figure out how to make the run results consistent, hence disable in cloud runs.
assert_binary_files_match(output_folder / f,
expected_folder / f)
def test_nii_load_zyx(test_output_dirs: OutputFolderForTests) -> None:
expected_shape = (44, 167, 167)
file_path = full_ml_test_data_path("patch_sampling/scan_small.nii.gz")
image: sitk.Image = sitk.ReadImage(str(file_path))
assert image.GetSize() == reverse_tuple_float3(expected_shape)
img = sitk.GetArrayFromImage(image)
assert img.shape == expected_shape
image_header = io_util.load_nifti_image(file_path)
assert image_header.image.shape == expected_shape
assert image_header.header.spacing is not None
np.testing.assert_allclose(image_header.header.spacing, (3.0, 1.0, 1.0), rtol=0.1)
@pytest.mark.parametrize("metadata", [None, PatientMetadata(patient_id="0")])
@pytest.mark.parametrize("image_channel", [None, known_nii_path, f"{good_h5_path}|volume|0", good_npy_path])
@pytest.mark.parametrize("ground_truth_channel",
[None, known_nii_path, f"{good_h5_path}|segmentation|0|1", good_npy_path])
@pytest.mark.parametrize("mask_channel", [None, known_nii_path, good_npy_path])
def test_load_images_from_dataset_source(
metadata: Optional[str],
image_channel: Optional[str],
ground_truth_channel: Optional[str],
mask_channel: Optional[str]) -> None:
"""
Test if images are loaded as expected from channels
"""
# metadata, image and GT channels must be present. Mask is optional
if None in [metadata, image_channel, ground_truth_channel]:
with pytest.raises(Exception):
def load_dataset_sources(
dataframe: pd.DataFrame,
local_dataset_root_folder: Path,
image_channels: List[str],
ground_truth_channels: List[str],
mask_channel: Optional[str],
allow_incomplete_labels: bool = False
) -> Dict[str, PatientDatasetSource]:
"""
Prepares a patient-to-images mapping from a dataframe read directly from a dataset CSV file.
The dataframe contains per-patient per-channel image information, relative to a root directory.
This method converts that into a per-patient dictionary, that contains absolute file paths
separated for for image channels, ground truth channels, and mask channels.
:param dataframe: A dataframe read directly from a dataset CSV file.
:param local_dataset_root_folder: The root folder that contains all images.
:param image_channels: The names of the image channels that should be used in the result.
:param ground_truth_channels: The names of the ground truth channels that should be used in the result.
:param mask_channel: The name of the mask channel that should be used in the result. This can be None.
:param allow_incomplete_labels: Boolean flag. If false, all ground truth files must be provided. If true, ground
truth files are optional. Default value is false.
:return: A dictionary mapping from an integer subject ID to a PatientDatasetSource.
"""
expected_headers = {
CSV_SUBJECT_HEADER, CSV_PATH_HEADER, CSV_CHANNEL_HEADER
}
# validate the csv file
actual_headers = list(dataframe)
if not expected_headers.issubset(actual_headers):
raise ValueError(
"The dataset CSV file should contain at least these columns: {}, but got: {}"
.format(expected_headers, actual_headers))
# Calculate unique data points, first, and last data point
unique_ids: List[str] = sorted(pd.unique(dataframe[CSV_SUBJECT_HEADER]))
if not local_dataset_root_folder.is_dir():
raise ValueError("The dataset root folder does not exist: {}".format(
local_dataset_root_folder))
def get_mask_channel_or_default() -> Optional[Path]:
if mask_channel is None:
return None
paths = get_paths_for_channel_ids(
channels=[mask_channel],
allow_incomplete_labels_flag=allow_incomplete_labels)
if len(paths) == 0:
return None
else:
return paths[0]
def get_paths_for_channel_ids(
channels: List[str],
allow_incomplete_labels_flag: bool) -> List[Optional[Path]]:
if len(set(channels)) < len(channels):
raise ValueError(f"ids have duplicated entries: {channels}")
rows = dataframe.loc[dataframe[CSV_SUBJECT_HEADER] == patient_id]
# converts channels to paths and makes second sanity check for channel data
paths, failed_channel_info = convert_channels_to_file_paths(
channels, rows, local_dataset_root_folder, patient_id,
allow_incomplete_labels_flag)
if failed_channel_info:
raise ValueError(failed_channel_info)
return paths
dataset_sources = {}
for patient_id in unique_ids:
metadata = PatientMetadata.from_dataframe(dataframe, patient_id)
dataset_sources[patient_id] = PatientDatasetSource(
metadata=metadata,
image_channels=get_paths_for_channel_ids(
channels=image_channels, # type: ignore
allow_incomplete_labels_flag=False),
mask_channel=get_mask_channel_or_default(),
ground_truth_channels=get_paths_for_channel_ids(
channels=ground_truth_channels, # type: ignore
allow_incomplete_labels_flag=allow_incomplete_labels),
allow_incomplete_labels=allow_incomplete_labels)
return dataset_sources
Checks if the shapes of the given tensors is equal, and the values are approximately equal, with a given
absolute tolerance.
"""
if isinstance(t2, list):
t2 = torch.tensor(t2)
assert t1.shape == t2.shape, "Shapes must match"
# Alternative is to use torch.allclose here, but that method also checks that datatypes match. This makes
# writing the test cases more cumbersome.
v1 = t1.flatten().tolist()
v2 = t2.flatten().tolist()
assert v1 == pytest.approx(
v2, abs=abs
), f"Tensor elements don't match with tolerance {abs}: {v1} != {v2}"
DummyPatientMetadata = PatientMetadata(patient_id=42)
def get_model_loader(
namespace: Optional[str] = None
) -> ModelConfigLoader[SegmentationModelBase]:
"""
Returns a ModelConfigLoader for segmentation models, with the given non-default namespace (if not None)
to search under.
"""
return ModelConfigLoader[SegmentationModelBase](
model_configs_namespace=namespace)
def get_default_azure_config() -> AzureConfig:
"""
expected = expected_file.read_bytes()
if actual == expected:
return
if actual_file.suffix == ".png" and expected_file.suffix == ".png":
actual_image = Image.open(actual_file)
expected_image = Image.open(expected_file)
actual_size = actual_image.size
expected_size = expected_image.size
assert actual_size == expected_size, f"Image sizes don't match: actual {actual_size}, expected {expected_size}"
assert np.allclose(
np.array(actual_image),
np.array(expected_image)), "Image pixel data does not match."
assert False, f"File contents does not match: len(actual)={len(actual)}, len(expected)={len(expected)}"
DummyPatientMetadata = PatientMetadata(patient_id='42')
def get_model_loader(
namespace: Optional[str] = None
) -> ModelConfigLoader[SegmentationModelBase]:
"""
Returns a ModelConfigLoader for segmentation models, with the given non-default namespace (if not None)
to search under.
"""
return ModelConfigLoader[SegmentationModelBase](
model_configs_namespace=namespace)
def get_default_azure_config() -> AzureConfig:
"""
def load_dataset_sources(
dataframe: pd.DataFrame, local_dataset_root_folder: Path,
image_channels: List[str], ground_truth_channels: List[str],
mask_channel: Optional[str]) -> Dict[int, PatientDatasetSource]:
"""
Prepares a patient-to-images mapping from a dataframe read directly from a dataset CSV file.
The dataframe contains per-patient per-channel image information, relative to a root directory.
This method converts that into a per-patient dictionary, that contains absolute file paths
separated for for image channels, ground truth channels, and mask channels.
:param dataframe: A dataframe read directly from a dataset CSV file.
:param local_dataset_root_folder: The root folder that contains all images.
:param image_channels: The names of the image channels that should be used in the result.
:param ground_truth_channels: The names of the ground truth channels that should be used in the result.
:param mask_channel: The name of the mask channel that should be used in the result. This can be None.
:return: A dictionary mapping from an integer subject ID to a PatientDatasetSource.
"""
expected_headers = {
CSV_SUBJECT_HEADER, CSV_PATH_HEADER, CSV_CHANNEL_HEADER
}
# validate the csv file
actual_headers = list(dataframe)
if not expected_headers.issubset(actual_headers):
raise ValueError(
"The dataset CSV file should contain at least these columns: {}, but got: {}"
.format(expected_headers, actual_headers))
# Calculate unique data points, first, and last data point
unique_ids = sorted(pd.unique(dataframe[CSV_SUBJECT_HEADER]))
if not local_dataset_root_folder.is_dir():
raise ValueError("The dataset root folder does not exist: {}".format(
local_dataset_root_folder))
def get_mask_channel_or_default() -> Optional[Path]:
if mask_channel is None:
return None
else:
return get_paths_for_channel_ids(channels=[mask_channel])[0]
def get_paths_for_channel_ids(channels: List[str]) -> List[Path]:
if len(set(channels)) < len(channels):
raise ValueError(f"ids have duplicated entries: {channels}")
paths: List[Path] = []
rows = dataframe.loc[dataframe[CSV_SUBJECT_HEADER] == patient_id]
for channel_id in channels:
row = rows.loc[rows[CSV_CHANNEL_HEADER] == channel_id]
if len(row) == 0:
raise ValueError(
f"Patient {patient_id} does not have channel '{channel_id}'"
)
elif len(row) > 1:
raise ValueError(
f"Patient {patient_id} has more than one entry for channel '{channel_id}'"
)
image_path = local_dataset_root_folder / row[
CSV_PATH_HEADER].values[0]
if not image_path.is_file():
raise ValueError(
f"The dataset references a file that does not exist: {image_path}"
)
paths.append(image_path)
return paths
dataset_sources = {}
for patient_id in unique_ids:
metadata = PatientMetadata.from_dataframe(dataframe, patient_id)
dataset_sources[patient_id] = PatientDatasetSource(
metadata=metadata,
image_channels=get_paths_for_channel_ids(
channels=image_channels), # type: ignore
mask_channel=get_mask_channel_or_default(),
ground_truth_channels=get_paths_for_channel_ids(
channels=ground_truth_channels) # type: ignore
)
return dataset_sources
