'Click and Grow' algorithm takes radiologist out of tumor segmentation

A tumor segmentation software dubbed single-click ensemble segmentation (SCES) is advancing the definition and delineation of a wide range of cancerous lesions past manual labor and into the realm of high-tech automation.

The concept behind the software is based on radiomics, or the high-throughput extraction and analysis of mine-able data from medical images.

“It must be high-throughput because we have to build databases that have thousands of images in them, not just tens or hundreds, and it has to be robust,” said Robert J. Gillies, PhD, chairman of the department of cancer imaging and metabolism and vice chair of radiology at Moffitt Cancer Center and Research Institute in Tampa, Fla.

The software runs seed imaging data, such as CT, through an algorithm of feature descriptors which provides all the characteristics of a potential tumor and subsequently churns out a 3D volumetric tumor segmentation detail with one click of the mouse.

Gillies developed the software with Philippe Lambin, MD, PhD, from the University Hospital Maastricht in Maastricht, The Netherlands. “This whole project got started based on a conversation that he and I had about four years ago. He’s a radiation oncologist and they have been developing similar databases for many years.”

The development of this algorithm and resulting software has been a massive undertaking spread out over several institutions and research groups from the Institute of Automation in Beijing, as well as Lambin’s group in The Netherlands and others in Germany and the UK. Continued crowd sourcing for improvements on the software are being developed based on a data cloud shared between Moffitt, Stanford, Harvard, Cornell and Columbia University.

“We are creating a collaborative environment in a cloud that would allow our developers to test their algorithms against a standard set of images and a documented evaluation matrix and see if they can do better to eliminate the need for human intervention, which is our goal,” said Gillies. 

The elimination of manual segmentation hints toward a new practice that is based on quantitative analysis, instead of human interpretation, which is not ideal. SCES is not a perfect solution, but it provides consistent, replicable results, which Gillies says is what really matters. 

“A lot of the arguments that people get into with segmentation is whether or not it is accurately representing the tumor itself,” he said  “We don’t get into that question. All we want is to get the same answer every time. If you do things manually and you allow the radiologist to delineate, then you consider that your gold standard and it’s a very low bar, indeed. Radiologists very rarely agree with each other. If you build the right software algorithm, you can get agreement more often than not--more than 95 percent--we get the same answer every time. I am not going to argue whether our segmentation is right or wrong. There is no right or wrong or gold standard. What we can say is, ‘are the results reproduceable,’ and yes, they are.”

This is where radiomics comes into play. The concept is that there are features in tumors in addition to volume that are predictive of outcome and response.

A single operator, for instance, enters original medical images in the SCES algorithm and a lesion is analyzed for the presence of more than 200 possible features. These include 13 descriptors of size, 14 ways to describe location, 12 descriptors of shape, three different ways to measure volume, 17 ways to describe histogram intensity and three different types of texture measurements. The algorithm also looks at co-occurrence matrices 17 ways, 30 different wavelets and a total of 125 different Law features.

SCES can be applied to lung cancer and cancers of the brain and breast, sarcoma, prostate and pancreatic cancers and melanoma. Multiple studies are ongoing. The entire library of features is able to catalogue lesions in a way that is comprehensive and entirely automated.

“We call the data mine-able because you can go back and find statistically significant relationships between these features and outcomes of response,” said Gillies.

In one study of lung cancer tumors from CT images, published October 2012 in the journal Pattern Recognition, 129 CT images were evaluated using the automated tumor segmentation software and results showed that the enterprise was “stable, accurate and automated.”

Image data are segmented by automatic application of the algorithm 20 times for every voxel and region of the suspected tumor. If the algorithm concludes at least 11 times that it is tumor, it is included in the segmented region of interest. If after those 20 cycles it is deemed to be tumor less than 11 times, it is not included in the resulting 3D volumetric segmentation map.

“We only accept something if it shows up more than 11 times,” said Gillies. “Unless it gets 11 votes, we don’t consider it to be a part of the tumor.”

There are few caveats so far in the software, but codevelopers keep working to better the standard algorithm and tweak the program. Clinical studies will continue to be conducted for a range of malignancies to improve the concept. “It does perform better than anything else out there, but that doesn’t mean it’s perfect.”

The software has been so popular that Gillies and his colleagues are being overrun with requests. The question of whether the software should be developed for commercial distribution hangs in the air.

“It was nothing we considered until about a month ago when we were just overwhelmed by requests to perform these operations,” said Gilles. “It is something we are considering, not because we are going to make money, but because it would provide a different source of income besides grants. With NIH grants being contracted, we have to start being creative about how we fund our research.”