Consensus of Ambiguity: Theory of Active Learning for Biomedical Image Analysis Applications

author: Scott Doyle, Laboratory for Computational Imaging and Bioinformatics, Rutgers, The State University of New Jersey
published: Oct. 14, 2010,   recorded: September 2010,   views: 3298
Categories

Slides

Related Open Educational Resources

Related content

Report a problem or upload files

If you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.
Lecture popularity: You need to login to cast your vote.
  Bibliography

Description

Supervised classifiers require manually labeled training samples to classify unlabeled objects. Active Learning (AL) can be used to selectively label only “ambiguous” samples, ensuring that each labeled sample is maximally informative. This is invaluable in applications where manual labeling is expensive, as in medical images where annotation of specific pathologies or anatomical structures is usually only possible by an expert physician. Existing AL methods use a single definition of ambiguity, but there can be significant variation among individual methods. In this paper we present a consensus of ambiguity (CoA) approach to AL, where only samples which are consistently labeled as ambiguous across multiple AL schemes are selected for annotation. CoA-based AL uses fewer samples than Random Learning (RL) while exploiting the variance between individual AL schemes to efficiently label training sets for classifier training. We use a consensus ratio to determine the variance between AL methods, and the CoA approach is used to train classifiers for three different medical image datasets: 100 prostate histopathology images, 18 prostate DCE-MRI patient studies, and 9,000 breast histopathology regions of interest from 2 patients. We use a Probabilistic Boosting Tree (PBT) to classify each dataset as either cancer or non-cancer (prostate), or high or low grade cancer (breast). Trained is done using CoA-based AL, and is evaluated in terms of accuracy and area under the receiver operating characteristic curve (AUC). CoA training yielded between 0.01-0.05% greater performance than RL for the same training set size; approximately 5-10 more samples were required for RL to match the performance of CoA, suggesting that CoA is a more efficient training strategy.

See Also:

Download slides icon Download slides: prib2010_doyle_cata_01.pdf (1.4 MB)


Help icon Streaming Video Help

Link this page

Would you like to put a link to this lecture on your homepage?
Go ahead! Copy the HTML snippet !

Write your own review or comment:

make sure you have javascript enabled or clear this field: