Evaluation of effects of jpeg2000 compression on a computer-aided detection system for prostate cancer on digitized histopathology

Scott Doyle, James Monaco, Anant Madabhushi, Stefan Lindholm, Patric Ljung, Lance Ladic, John Tomaszewski, Michael D. Feldman
IEEE International Symposium on Biomedical Imaging, page 1313--1316 - 2010
Download the publication : Doyle et al. - 2010 - JPEG2000 CAD.pdf [1.7Mo]  
A single digital pathology image can occupy over 10 gigabytes of hard disk space, rendering it difficult to store, analyze, and transmit. Though image compression provides a means of reducing the storage requirement, its effects on computer-aided diagnosis (CAD) and pathologist performance are not yet clear. In this work we assess the impact of compression on the ability of a CAD system to detect carcinoma of the prostate (CaP) on histological sections. The CAD algorithm proceeds as follows: Glands in the tissue are segmented using a region-growing algorithm, and the size of each gland is extracted. A Markov prior (specifically, a probabilistic pairwise Markov model) is employed to encourage nearby glands to share the same class (i.e. cancerous or non-cancerous). Finally, cancerous glands are aggregated into continuous regions using a distancehull algorithm. We trained the CAD system on 28 images of wholemount histology (WMH) and evaluated performance on 12 images compressed at 14 different compression ratios (a total of 168 experiments) using JPEG2000. Algorithm performance (measured using the under the receiver operating characteristic curves) remains relatively constant for compression ratios up to1 :256, beyond which performance degrades precipitously. For completeness we also have an expert pathologist view a randomly-selected set of compressed images from one of the whole mount studies and assign a confidence measure as to their diagnostic fidelity. Pathologist confidence declined with increasing compression ratio as the information necessary to diagnose the sample was lost, dropping from 100% confidence at ratio 1:64 to 0% at ratio 1:8192.

Images and movies

 

BibTex references

@inproceedings{DMMLLLTF10,
  author       = {Doyle, Scott and Monaco, James and Madabhushi, Anant and Lindholm, Stefan and Ljung, Patric and Ladic, Lance and Tomaszewski, John and Feldman, Michael D.},
  title        = {{Evaluation of effects of jpeg2000 compression on a computer-aided detection system for prostate cancer on digitized histopathology}},
  booktitle    = {IEEE International Symposium on Biomedical Imaging},
  pages        = {1313--1316},
  year         = {2010}
}

Author publication list