CAD developer targets false positives on RSNA floor

RSNA newcomer Parascript takes aim this week at false positives in computer-aided detection, hoping to convince other vendors that its proprietary image analysis software, proven in fields outside radiology, can do a better job than other mammography CAD systems.

In demonstrations on the exhibit floor, the company is targeting the biggest complaint of radiologists who use mammography CAD, as false positives waste time in an already time-precious environment. The more false positives, the more time wasted, as they go back to images to rule out suspicious lesions. This makes CAD one of the few technological innovations that slows rather than speeds patient throughput.

AccuDetect has the potential to cut false positives by 60% to 80% below those typically encountered in mammography CAD, according to Yuri Prizemin, Parascript's director of product marketing. Prizemin is careful, however, in choosing his words.

"Parascript is not making any direct claims," he said. "Instead we are saying that we 'anticipate' that existing false-positive rates can be reduced (by this much) with the use of AccuDetect algorithms."

Actual claims about its performance, according to Prizemin, must wait until AccuDetect is actually integrated with mammography systems. The company is grooming AccuDetect for sale to OEM customers. The software is compatible with both film-based and digital mammography exams, according to Prizemin.

Underlying AccuDetect algorithms is a dozen years' experience in software that digitizes and analyzes printed information for the U.S. Postal Service and corporations including Lockheed Martin, NCR, Siemens, and Unisys.

"Medical imaging is a natural extension of our technology into areas that face significant challenges and require further technology advancements," Prizemin said. "We believe the medical imaging breast cancer detection market is facing the challenge of high false-positive rates and are confident our proprietary image analysis technology can address these issues."

A differentiator between AccuDetect and other such software is the way its algorithms mark suspected lesions without obscuring the region of interest, he said.

"Marking suspected lesions may often interfere with the region of interest or the area identified as suspect, prohibiting radiologists from analyzing the area more closely," Prizemin said. "Each CAD application has a different way of dealing with this issue."

Parascript's approach is to adjust settings so marked areas can be either de-emphasized or temporarily deleted. The unobtrusive markings enable more accurate interpretation of mammograms and increased detection of calcifications and masses.

Adding to the power of AccuDetect is the ability of algorithms to learn. With increasing experience, the software learns to better utilize the programmed knowledge.

After training is complete, the system can automatically locate and recognize similar but not necessarily identical lesions and cases. This training is an iterative process and includes testing and sometimes changing the system structure and its parameters.

The software also makes use of "voting methods" to boost sensitivity and lower false-positive rates. Voting takes the output from two or more algorithms and compares the results.

"Different algorithms can give multiple answers, such as confidence levels, types, locations, and dimensions of breast lesions, and these output results are used in voting," Prizemin said.

Whether their analyses will translate into substantially fewer false positives remains to be proven, but Parascript's decision to enter this field underscores that there is at least room for improvement in the software now available.