Digital Pathology -- Segmentation of Nuclei in Images

Organized by cpm.organizing.committee - Current server time: Sept. 26, 2017, 8:56 a.m. UTC

First phase

Training
June 1, 2017, midnight UTC

End

Competition Ends
July 31, 2017, 11:59 p.m. UTC

Overview

Please note the test phase will start on July 20 and will end on July 31.

Grading and diagnosis of tumors in cancer patients have traditionally been done by examination of tissue specimens under a powerful microscope by expert pathologists. While this process continues to be widely applied in clinical settings, it is not scalable to translational and clinical research studies involving hundreds or thousands of tissue specimens. State-of-the-art digitizing microscopy instruments are capable of capturing high-resolution images of whole slide tissue specimens rapidly. Computer aided segmentation and classification has the potential to improve the tumor diagnosis and grading process as well as to enable quantitative studies of the mechanisms underlying disease onset and progression.

The objective of the Digital Pathology Challenge is to evaluate and compare classification and segmentation algorithms and to encourage the biomedical imaging community to design and implement more accurate and efficient algorithms. This year’s challenge targets algorithms for whole slide tissue images obtained from patients with non small cell lung cancer (NSCLC), head and neck squamous cell carcinoma (HNSCC), glioblastoma multiforme (GBM), and lower grade glioma (LGG) tumors. These cancer types are complex and deadly diseases accounting for a large number of diagnostic patient deaths in spite of application of various treatment strategies.

The goal of this sub-challenge is to evaluate the performance of algorithms for detection and segmentation of nuclear material in a tissue image. Participants are asked to detect and segment all the nuclear material in a given set of image tiles extracted from whole slide tissue images. This sub-challenge uses image tiles from whole slide tissue images to reduce computational and memory requirements.

The image tiles are rectangular regions extracted from a set of GBM, LGG, HNSCC, and NSCLC whole slide tissue images. Nuclei in each image tile in the training set are manually segmented. The training set consists of 32 image tiles, 8 tiles from each cancer type. The image tiles are available as PNG images. The segmentation result for each tile is available as a labeled text mask file. A labeled mask file is an array of integer values. The array has the same resolution as the image tile. Each array element stores a value between 0 and N, where N is the number of segmented nuclei. A value of 0 in an array element means the corresponding pixel in the image tile is not part of a nucleus. A non-zero value means the corresponding pixel in the image tile is part of a nucleus. All the pixels that belong to the same nucleus are labeled with the same value.

Evaluation

Please note the test phase will start on July 20 and will end on July 31.

Each participant will be required to submit an extended abstract (up to 4 pages) that describes their algorithm at the end of the training phase (update: Please email your abstracts by Monday, July 24). You may submit your abstract to tahsin.kurc@stonybrook.edu .

Each participant will be required at the end of the test phase to submit a zip file containing mask files. The mask files are text files with the following format:

width height     // N M
pixel_label_id  // pixel (0,0)
pixel_label_id  // pixel (1,0)
pixel_label_id  // pixel (2,0)
…
pixel_label_id  // pixel (N-1,M-1)

For example, assume two nuclei were segmented in a tile of 5x4 pixels:

00000
01100
11200
02220

The mask file would have the following content:

5 4
0
0
0
0
0
0
1
1
0
0
1
1
2
0
0
0
2
2
2
0

The zip file should have the following folder structure: 
gbm/
hnsc/
lgg/
lung/

Each folder should contain the mask files generated from the respective test images. The filename of each mask file must be the same as the image file prefix followed by a “_mask” suffix. For example, the mask file for image1.png must be image1_mask.txt.

The scoring for this sub-challenge will be done using two variants of the DICE coefficient: Traditional Dice coefficient (DICE_1) to measure the overall overlapping between the reference/human segmentation and the participant segmentation.

An “Ensemble Dice” (DICE_2) to capture mismatch in the way the segmentation regions are split, while the overall region may be very similar. The two DICE coefficients will be computed for each image tile in the test dataset. The score for the image tile will be the average of the two dice coefficients. The score for the entire test dataset will be the average of the scores for the image tiles.

Terms and Conditions

Please note the test phase will start on July 20 and will end on July 31.

This page enumerated the terms and conditions of the competition.

No submissions have been made public!

Training

Start: June 1, 2017, midnight

Test

Start: July 20, 2017, midnight

You must be logged in to participate in competitions.

Sign In
No submissions have been made public!
No submissions have been made public!