Computer assistance targets gray areas in breast ultrasound

New computer-aided diagnosis programs are promising to help radiologists improve their performance in the highly challenging area of breast ultrasound. Up to one woman in five who gets a screening mammogram also undergoes an ultrasound study. Among other uses, breast ultrasound helps in the diagnosis of women who have dense breasts or are pregnant.

Ultrasound's potential as a screening tool is promising, but it is still under investigation in large-scale trials. If breast ultrasound becomes accepted for screening, CAD products could be adapted with additional clinical validation and research.

"We will talk about using and extending our technology for reducing false negatives," said Michael Galperin, CEO of Almen Laboratories, maker of a breast ultrasound computer-aided tool for diagnostic support. "Right now, we are working only on reduction of biopsies, rather than catching missed cancers."

Though ultrasound has considerable potential, it also has well-known limitations. Technique is difficult to teach and perform, and differentiating between benign and malignant lesions can be very challenging. The addition of ultrasound could also lead to unnecessary biopsies of benign lesions.

An excess of biopsies is already a big clinical and economic problem. Every year, 1.7 million women have core needle or surgical biopsies, and the majority-75% to 80%-have a benign result. These biopsies cost more than $4 billion, in addition to the personal toll.

"A breast biopsy of a benign lesion is anything but a benign procedure," Galperin said.

Almen Laboratories expects to have FDA 510(k) clearance of its CAD system by the end of summer. The system, developed in conjunction with researchers at the University of California, San Diego, helps radiologists identify objects and image features in regions of interest. The computer compares a breast mass with a database of patient images with biopsy-proven findings and estimates a level of suspicion score based on similarity to other cases. Lesions are scored according to the American College of Radiology BI-RADS lexicon.

BIOPSY THRESHOLDS

This tool could prove most helpful in difficult cases. Radiologists automatically order a biopsy of mammographic lesions categorized as BI-RADS 4 or 5, while dismissing lesions that are obviously benign, Galperin said. It's the lesions in the BI-RADS 2 or 3 categories that prove challenging.

"Radiologists have different thresholds for biopsy. For some, if one or two malignant categories from BI-RADS are present, they don't hesitate to biopsy. Others would like to look more closely at the case, and possibly seek help from a computer, before opting for biopsy," he said.

Greater willingness to biopsy has implications down the line for the patient.

"If that lesion is a simple cyst or fibroadenoma, it cannot be diagnosed after biopsy, even if it is benign, because of internal scarring," Galperin said.

This consequence is particularly problematic for women who move without taking their case histories with them and/or those who are evaluated by a new radiologist. The scarring will look suspicious on subsequent mammograms and ultrasound, possibly triggering additional biopsies.

"There is a very high danger of repeat biopsy on a benign lesion," he said.

Galperin said the computer assistance should boost radiologists' confidence in determining whether lesions in that gray area are benign or malignant. He suggests that lesions of a lower level of suspicion such as cysts may be ruled out as candidates for biopsy with a higher degree of confidence when image interpretation is made with support of the computer-aided imaging system.

The company has completed four clinical studies using the CAD system. The most recent was a two-year retrospective study of 332 lesions that were biopsy-proven or, in the cases where biopsy was not required, followed for two years. Of the 332 cases, 22% were malignant and 78% were benign.

In comparison with previous trials with earlier versions of the software, the latest study showed more promising results:

88% sensitivity;

97% specificity;

90% positive predictive value; and

96% negative predictive value.

Performance of the software alone was compared with that of four subspecialty radiologists, all of whom had more than 15 years' experience in breast imaging. The program's specificity of 97% compared favorably with that of the doctors, whose scores ranged from 58% to 91%. Two of the doctors scored better in sensitivity (92% and 93%), however, compared with CAD alone. The software's diagnostic efficacy of 95% was higher than that of all four doctors, who scored from 68% to 83% in diagnostic efficacy.

RELEVANT FEATURES

Cedara Software of Ontario has also reported positive experiences. The company received FDA 510(k) clearance in 2005 for B-CAD, a computer-aided program designed to enhance radiologists' performance in analysis of breast ultrasound. B-CAD is based on BI-RADS and incorporates the ACR's BI-RADS Ultrasound Lexicon Classification Form. The system analyzes, segments, and classifies shape and orientation characteristics in regions of interest selected by the user.

A new version of the system that covers a range of other sonographic features has been developed and is under review by the FDA. Results of two retrospective studies with the new B-CAD version were presented at the American Institute of Ultrasound in Medicine meeting in March by Dr. A. Thomas Stavros, a radiologist at the Sally Jobe Breast Center in Colorado, and sonographer Cynthia Rapp of the same institution.

One study consisted of 30 biopsy-confirmed single-view images of fibroadenomas in 21 cases. B-CAD identified the BI-RADS features of the fibroadenomas consistent with an expert radiologist. It also categorized the majority of fibroadenomas as BI-RADS 1 to 3.

The second study evaluated B-CAD's ability to analyze and classify breast lesions. Researchers used single-view images of 74 breast lesions in 47 cases: 40 were malignant, 34 benign. The agreement between an expert radiologist and the CAD system in identifying BI-RADS features ranged up to 90% (scores varied depending on the sonographic feature tested). Positive and negative predictive values were up to 88% and 85%, respectively.

Although the ACR lexicon has standardized ultrasound reporting, radiologists may still not be looking for all of the relevant features, said Sabrina Cannistraro, Cedara's product manager for women's health.

"There is so much to look for-it's a pretty extensive list. B-CAD automatically highlights those features. The list is directly in front of you and serves as a reminder of what to look for," she said.

In the U.S., technologists typically perform breast ultrasound scanning, while the radiologist reviews and interprets the images. Presently, the B-CAD system can be used at the review station. As the company expands outside the U.S., it is looking to incorporate use of the tool during scanning as well.

Emily Hayes is feature editor for Diagnostic Imaging.