import xml.etree.ElementTree as ET
from functools import cached_property
from pathlib import Path
from typing import Dict, Optional, Sequence, Union
import albumentations
import numpy as np
import tifffile
from PIL import Image
from skimage.draw import polygon
from skimage.util import img_as_float32
from ..base import MaskDataset
[docs]class ComputationalPathology(MaskDataset):
"""A Dataset and a Technique for Generalized Nuclear Segmentation for
Computational Pathology [1]_
Parameters
----------
root_dir : str
Path to root directory
output : {'both', 'image', 'mask'}, default: 'both'
Change outputs. 'both' returns {'image': image, 'mask': mask}.
transforms : albumentations.Compose, optional
An instance of Compose (albumentations pkg) that defines augmentation in
sequence.
num_samples : int, optional
Useful when ```transforms``` is set. Define the total length of the
dataset. If it is set, it overwrites ``__len__``.
grayscale : bool, default: False
Convert images to grayscale
grayscale_mode : {'cv2', 'equal', Sequence[float]}, default: 'cv2'
How to convert to grayscale. If set to 'cv2', it follows opencv
implementation. Else if set to 'equal', it sums up values along channel
axis, then divides it by the number of expected channels.
mask_tif : bool, default: False
Instead of parsing every xml file to reconstruct mask image arrays, use
pre-drawn mask tif files which should reside in the same folder as
annotation xml files.
Notes
-----
- Resolution of all images is (1000,1000)
- gt is converted from annotation recorded in xml format
- gt has dtype of torch.float64, converted from numpy.uint16, and it has
value 'num_objects' * 255 because it is base-transformed
- The origianl dataset provides annotation in xml format, which takes
long time to parse and to reconstruct mask images dynamically during
training. Drawing masks beforehand makes training much faster. Use
``mask_tif`` in that case.
- When ``augmenters`` is provided, set the ``num_samples`` argument
30x1000x1000 -> 16x30=480 patches. Thus, the default ``num_samples=720``
(x1.5)
- dtype of 'gt' is int16. However, to make batching easier, it will be
casted to float32
- Be careful about types of augmenters; avoid interpolation
References
----------
.. [1] N. Kumar, R. Verma, S. Sharma, S. Bhargava, A. Vahadane, and A.
Sethi, “A Dataset and a Technique for Generalized Nuclear Segmentation
for Computational Pathology,” IEEE Transactions on Medical Imaging, vol.
36, no. 7, pp. 1550–1560, Jul. 2017, doi: 10.1109/TMI.2017.2677499.
See Also
--------
MaskDataset : Super class
Dataset : Base class
DatasetInterface : Interface
"""
# Dataset's acronym
acronym = 'ComPath'
# Hard code resolution to parse annotation (.xml)
_resolution = (1000, 1000)
def __init__(
self,
root_dir: str,
*,
output: str = 'both',
transforms: Optional[albumentations.Compose] = None,
num_samples: Optional[int] = None,
grayscale: bool = False,
grayscale_mode: Union[str, Sequence[float]] = 'cv2',
# specific to this dataset
mask_tif: bool = False,
**kwargs
):
self._root_dir = root_dir
self._output = output
self._transforms = transforms
self._num_samples = num_samples
self._grayscale = grayscale
self._grayscale_mode = grayscale_mode
# specific to this dataset
self.mask_tif = mask_tif
[docs] def get_image(self, p: Path) -> np.ndarray:
img = Image.open(p)
if img.mode == 'RGBA':
img = img.convert(mode='RGB')
return img_as_float32(np.asarray(img))
[docs] def get_mask(self, p: Path) -> np.ndarray:
if self.mask_tif:
if p.suffix == '.xml':
raise ValueError(
"Use `save_xml_to_tif()` then set `mask_tif` to True"
)
mask = tifffile.imread(p)
return mask.astype(np.int16)
# Parse xml
mask = self._parse_xml_to_array(p)
return mask
@classmethod
def _parse_xml_to_array(cls, f_anno) -> np.ndarray:
"""This dataset provides annotation in .xml format
Consider pre-generating mask image using ``save_xml_to_tif()``
"""
tree = ET.parse(f_anno)
root = tree.getroot()
rr = []
cc = []
for region in root.iter('Region'):
r = []
c = []
# print(region.attrib)
# print(region.find('Vertices'))
if (vertices := region.find('Vertices')) is not None:
for v in vertices:
# print(v.attrib)
r.append(v.attrib['Y'])
c.append(v.attrib['X'])
rr.append(np.array(r, dtype=np.float16))
cc.append(np.array(c, dtype=np.float16))
# X, Y = anno['X'], anno['Y']
mask = np.zeros(cls._resolution, dtype=np.int16)
for i, (x, y) in enumerate(zip(cc, rr), 1):
r, c = polygon(y, x, shape=cls._resolution)
if len(rr) == 0 and len(cc) == 0:
continue
mask[r, c] = i
return mask
# return {'X': cc, 'Y': rr}
[docs] def save_xml_to_tif(self):
"""Parse .xml to mask and write it as tiff file
Having masks in images is much faster than parsing .xml for each call.
This func iterates through ``anno_dict``, parse and save each in .tif
format in the same annotation directory. Re-initiate an instance with
``mask_tif`` argument to load them.
"""
if self.output not in ['mask', 'both']:
raise ValueError("Set output either to 'mask' or 'both'")
for i, p in self.anno_dict.items():
mask = self._parse_xml_to_array(p)
fname = p.with_suffix('.tif')
tifffile.imwrite(
fname,
data=mask,
compression='zlib'
)
print(f"[{i}/{len(self.anno_dict) - 1}] Wrote '{fname}'")
@cached_property
def file_list(self) -> list:
root_dir = self.root_dir
parent = 'Tissue images'
file_list = sorted(root_dir.glob(f'{parent}/*.png'))
return file_list
@cached_property
def anno_dict(self) -> Dict[int, Path]:
root_dir = self.root_dir
parent = 'Annotations'
ext = 'xml'
if self.mask_tif:
ext = 'tif'
anno_dict = dict((k, v) for k, v in enumerate(
sorted(root_dir.glob(f'{parent}/*.{ext}'))
))
return anno_dict