Content

A collection of datasets converted into COCO segmentation format.

Preprocessing:

• Resized few images

• Tiled some images with lot of annotations to fit in memory

• Extracted masks when only outlines were available

• This is done by finding contours

Folder hierarchy

DATASETS = {
    'nuclei_stage1_train': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/stage_1_train',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/stage1_train.json'
    },
    'nuclei_stage_1_local_train_split': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/stage_1_train',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/stage_1_local_train_split.json'
    },
    'nuclei_stage_1_local_val_split': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/stage_1_train',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/stage_1_local_val_split.json'
    },
    'nuclei_stage_1_test': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/stage_1_test',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/stage_1_test.json'
    },
    'nuclei_stage_2_test': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/stage_2_test',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/stage_2_test.json'
    },
    'cluster_nuclei': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/cluster_nuclei',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/cluster_nuclei.json'
    },
    'BBBC007': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/BBBC007',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/BBBC007.json'
    },
    'BBBC006': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/BBBC006',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/BBBC006.json'
    },
    'BBBC018': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/BBBC018',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/BBBC018.json'
    },
    'BBBC020': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/BBBC020',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/BBBC020.json'
    },
    'nucleisegmentationbenchmark': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/nucleisegmentationbenchmark',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/nucleisegmentationbenchmark.json'
    },
    '2009_ISBI_2DNuclei': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/2009_ISBI_2DNuclei',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/2009_ISBI_2DNuclei.json'
    },
    'nuclei_partial_annotations': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/nuclei_partial_annotations',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/nuclei_partial_annotations.json'
    },
    'TNBC_NucleiSegmentation': {
        IM_DIR:
            _DATA_DIR + '/Nuclei/TNBC_NucleiSegmentation',
        ANN_FN:
            _DATA_DIR + '/Nuclei/annotations/TNBC_NucleiSegmentation.json'
    },
}

Example usage:

import json
from pathlib import Path
import numpy as np
from PIL import Image
from pycocotools import mask as mask_util

ROOT_DIR = Path('/media/gangadhar/DataSSD1TB/ROOT_DATA_DIR/')
DATASET_WORKING_DIR = ROOT_DIR / 'Nuclei'

annotations_file = DATASET_WORKING_DIR / 'annotations/stage1_train.json'

COCO = json.load(open(annotations_file.as_posix()))

image_metadata = COCO['images'][0]
print image_metadata

# {u'file_name': u'4ca5081854df7bbcaa4934fcf34318f82733a0f8c05b942c2265eea75419d62f.jpg',
#  u'height': 256,
#  u'id': 0,
#  u'nuclei_class': u'purple_purple_320_256_sparce',
#  u'width': 320}


def get_masks(im_metadata):
    image_annotations = []
    for annotation in COCO['annotations']:
        if annotation['image_id'] == im_metadata['id']:
            image_annotations.append(annotation)

    segments = [annotation['segmentation'] for annotation in image_annotations]
    masks = mask_util.decode(segments)
    return masks


masks = get_masks(image_metadata)

print masks.shape
# (256, 320, 37)


def show(i):
    i = np.asarray(i, np.float)
    m,M = i.min(), i.max()
    I = np.asarray((i - m) / (M - m + 0.000001) * 255, np.uint8)
    Image.fromarray(I).show()


show(np.sum(masks, -1))
# this should show an image with all masks