Laboratory for Knowledge Inference in
Medical Image Analysis

  Home About us Research The Team Projects Code and Data  

Inter-Observer Variability in Image Segmentation:
500 Synthetic Prostate Ultrasound Images + 20 Segments/Image

This dataset was proposed in the following papers:

  1. Fast Barcode Retrieval for Consensus Contouring
    H.R.Tizhoosh , G.J.Czarnota, arXiv:1709.10197v1 [cs.CV] 28 Sep 2017 

  2. Anatomy-Aware Measurement of Segmentation Accuracy
    H.R.Tizhoosh, A.A. Othman

TRUS images of prostates may be used to both diagnose and treat prostate diseases such as cancer. Starting with a set of prostate shapes P1, P2,..., Pm, we created random segments Gi through combinations of those priors, adding noise along with random translations and rotations, and we distorted the results with speckle noise and shadow patterns. Each image Ii is thus created from its gold Gi. Consequently, we can simulate k user delineations Si1, Si2,..., Sik by manipulating Gi via scaling, rotation, and morphological changes, and we can simulate edits by running active contours with variable user-simulating parameters. The variability of user delineations was simulated according to several factors: error probability ([0,0.05]), anatomical difficulty (=0.2 out of [0, 1]), and the scaling factor for morphology (form 1x1 to 21x21). The user was modelled according to the level of experience (a random number from (0,1]), the user's attention (a random number from [0,1]), and the user's tendencies in terms of the segment size (a random number from [-1,1]), whether tending to draw contours that are relatively small (->-1) or large (->+1).

We generated 500 images from their corresponding gold-standard images. Furthermore, we generated 20 different segments for each image, assuming that there were 20 users. The figure below shows an example of the gold segments and simulated user contours. The variability, coupled with the gold segment, is what is needed to validate our approach.

Synthetic Prostate

Figure. Sample image shows gold segments and consensus contours (left). The users have drawn the contour differently (middle, with the gold contour superimposed). The inverted region (middle) is magnified (right) to emphasize details of the variability.


:: Download the paper 1: Fast Barcode Retrieval...

:: Download the paper 2: Anatomy-Aware Measurement...

:: Download the dataset [11,500 images (500 images each with 20 segments and consensus); 38 MB]

:: Our Sponsors




© KIMIA Lab, University of Waterloo, Canada, 2013-2017