Fig. 1. Overview of the proposed nucleus segmentation.
Matlab implementation
BBBC/ISBI test dataset (partial)
BBBC full dataset(external link)
ISBI full dataset(external link)
Test instruction
1. Download source codes and unzip it
2. Download the test data and unzip it to the "/data" in the source code folder. Create the "data" folder if necessary.
3. Launch the Matlab and run Nucleus_segmentation.m with a parameter containing the list of filesname(s) to test.
For example,
> filelist = ls('data/BBBC');
> Nucleus_segmentation('data/BBBC/', filelist);
4. To test the full dataset, download the full dataset from the external site, and follow the instruction 1-3.
Dataset
◾
BBBC006v1 in Broad Bioimage Benchmark Collection (BBBC) dataset
The BBBC dataset consists of 768 microscopy images and
corresponding ground-truth images, produced by experts. The
resolution of each image has 696×520, and pixels are 16-bit
grayscale. In this dataset, the contrast between the background
and nuclei varies from image to image: in our image the nuclei
may have high intensity; in another their intensity might be
low. Ten images were excluded from our experiments because
their ground truth images did not contain any labeled nuclei.
◾ 2009 IEEE International Symposium on Biomedical Imaging (ISBI) dataset
The ISBI dataset consists of 48 microscopy images and
their ground-truth images segmented by hand. The resolution
of each image is 1349×1030 and pixels are 16-bit RGB.
To allow these two datasets to be analyzed in a similar way,
the RGB data was converted to grayscale, and the image
size was reduced to 365×279. The sizes of nuclei in these reduced images were similar to the sizes of nuclei in the
BBBC dataset. In this IBSI dataset, the contrast between
background and nuclei varies between images, like the BBBC
data, but this second dataset is more challenging because the
nuclei tend to exhibit bumpy surface and irregular intensities.
Moreover, some nuclei have shapes which are neither circular
nor elliptical.
Experimental results
Each nucleus was randomly assigned a unique color and a number after segmentation by below methods.
(a) Original image
(b) Ground truth image
(c) Local+Global+Fusion+Fission+Smooth version of our method
(d) GC
(e) MINS
(f) OTSU
(g) ilastik
(h) ER
BBBC dataset
ISBI dataset
Note: This page is for review purpose only. Please DO NOT distribute or share the link without authors permission.