Encrypted login | home

Program Information

Automatic Organ at Risk Delineation with Machine Learning Techniques

no image available
G Bernard

G Bernard*, M Verleysen , J Lee , Universite catholique de Louvain, Brussels

Presentations

SU-C-18A-3 Sunday 1:00PM - 1:55PM Room: 18A

Purpose:
Manual delineation of organs at risk (OARs) on CT images consumes much time. Automatic segmentation methods like atlases partly address this issue. However, atlases depend on deformable registration quality. This work proposes an atlas-like method that relies on machine learning techniques instead of registration.

Methods:
First, a watershed algorithm segments filtered CT images into superpixels (images patches with similar intensity pixels). Next, two kinds of superpixel features are computed: intrinsic ones (known at all times, like superpixel size, position, and mean intensity) and extrinsic ones (to be inferred from partial delineation results, like the distances to other organs). To build the atlas, a binary classifier is associated with each organ. Training optimizes the classifiers' parameters as well as their sequence, to make the most useful extrinsic features available as soon as possible. After training, the sequence of binary classifiers can process any new image, tagging all superpixels incrementally with an OAR label.

The method was applied to 2D CT images of 49 breast-cancer patients (axial slice passing through the 7th thoracic vertebra). The balanced classification rate (BCR) measures the method's accuracy, by giving the percentage of correctly classified pixels per label.

Results:
The proposed incremental method (BCR = 94%) is compared to two similar classification procedures, with either no extrinsic features (blind, BCR = 84%) or all of them known beforehand (cheating oracle, BCR = 98%). A preliminary comparison with a registration-based atlas on synthetic data led to 97% (registration) and 98% (proposed).

Conclusion:
This abstract demonstrates the feasibility of atlas-like OAR delineation based on machine learning techniques instead of deformable registration. The proposed method relies on incremental classification (partial classification allows additional highly informative features to be inferred). Involving no elastic deformation, delineation can be easily corrected if needed, just by changing the erroneous superpixel labels.

Funding Support, Disclosures, and Conflict of Interest: John Aldo Lee is a Research Associate with the Belgian fund of scientific research F.R.S-FNRS.


Contact Email: