PICSL alumni Nick Tustison and colleagues Max Wintermark, Chris Durst, and Brian Avants finished in first place in the Multimodal Brain Tumor Segmentation challenge at the MICCAI conference (BRATS 2013):
Given the success of random forest approaches for segmentation, particularly for the BRATS 2012 tumor segmentation challenge, we implemented a variant framework for our own research. The innovation of our methodology and implementation is characterized by the following four-fold contribution: 1) generation of novel feature images in addition to what has been previously reported which significantly en- hances classification, 2) concatenated application of random forest mod- els for improved performance, 3) the use of ANTsR (a packaging of the ANTs library plus additional analysis tools for the R statistical project) for direct access to robust random forest functionality with parallelization, and 4) public availability of all scripts to recreate the leave-one-out evaluation study performed with the provided training data.
The full description of the algorithm can be found in the BRATS proceedings, which includes links to the tools and scripts.
The evaluation results, public data, and more information on the challenge is available at the BRATS site.