Computer Aided Detection of Masses in Mammograms

From UMCN Radiology Research

General project information

• Title: Computer Aided Detection of Masses in Mammograms
• Acronym :
• Time frame: 2006 - 2010
• Funded by KWF (project: Earlier detection of breast cancer by computer assisted decision making in screening)

Project outline

Introduction

Breast cancer screening programs have been established as an effective way to reduce mortality from breast cancer. It is well known that the effectiveness of these programs strongly depends on high acceptance by the population and on high quality of the screening procedure. While technical quality assurance has received a lot of attention to guarantee optimal mammographic image quality, the quality of mammographic interpretation now seems to be the weakest link in the process. Review studies have shown that observer errors are frequent in breast cancer screening. On the basis of these studies it is estimated that 20 to 30 percent of the cancers could be detected in an earlier screening round without increasing the recall rate to an unacceptable level.

To reduce the number of false negatives in screening we develop computer aided diagnosis (CAD) methods. Currently, these systems are primarily designed to provide prompts at suspicious areas in mammograms to focus the attention of radiologists. Typically, these prompting systems operate at a high sensitivity, but their specificity is only modest. The idea behind these CAD systems is that when a radiologist carefully inspects mammographic areas that are prompted the risk of overlooking a significant abnormality is minimized. More recently, research is also focussed at helping radiologists to interpret detected lesions in order to improve the decision making process.

Current research projects include automated detection and characterization of masses and microcalcification clusters, the two major mammographic signs of breast cancer. Large annotated databases of digitized mammograms have been collected, with known abnormalities to facilitate development. Observer studies are carried out to measure the performance of radiologists with and without CAD.

Some examples are outlined below.

Detection of Masses

n the majority of cases missed by screening a mass is involved, which is the most important mammographic sign for detection of invasive breast cancer. Masses may be hard to detect as they are easily obscured by the fibroglandular tissue, especially when the tumor is still small. It appears that most malignant densities have irregular shapes and that they are frequently surrounded by a radiating pattern of linear spicules. Sometimes the central density is faint or absent. In those cases the stellate pattern of spicules is the most important sign. Pixelwise classification based on local features also forms the basis of a method for detection of stellate distortion that we developed. In this method detection is performed by statistical analysis of a map of line-based pixel orientations. The idea is that if a strong increase of pixels pointing to a given region is found this region is suspicious, especially if such an increase is found in many directions. No attempt is being made to explicitly identify spicules. The calculations are performed at a grid of sites inside the imaged breast area. Afterwards, neighboring sites with a high level of suspiciousness are linked up to form regions. Before calculation of the pixel orientation map, the breast tissue is segmented from the background, and in the oblique views the pectoral muscle is marked. An example of this segmentation, which is performed fully automatic, is shown in Figure 3. It used to perform a filtering step in which the pectoral muscle is removed and in which the decrease of image intensities near the breast skin line is compensated for. The processed image is used as input to a bank of filters for detection of bright areas at a range of spatial scales. When such a bright area is present, its scale is used to set the size of the surroundings to be analyzed for detection of spiculation. In case a small density is present the program analyzes a smaller region than in case of a larger density.

Pixel orientations are estimated using a new method based on Gaussian scale-space theory. If a line-like structure is present at a given pixel this method provides an accurate estimate of its orientation, whereas in other cases the image noise generates some random output. The orientation estimates are used to construct two operators which respond to radial patterns of straight lines. The first one is defined to measure the total number of pixels with directions pointing to a test area. If this number is significantly more than would be expected on the basis of randomness, the area may be suspicious. However, if an increase of the number of pixels oriented towards a region is found in a few directions only, it isn't very likely that the site being evaluated belongs to the center of a stellate pattern. On the other hand, if evidence for spicules is found in many directions this should increase the likelihood of a stellate structure being present. To represent this property, a second operator is constructed, measuring the uniformity of the orientation map. The two features defined are combined to form an image representing the degree of suspiciousness at each pixel. Figure 3d shows an example, where bright spot is at the location of a histologically verified malign stellate distortion.

Observer Studies

We investigated the use of a computer aided diagnosis (CAD) system designed for detection of mammographic masses as an aid to improve interpretation, and to compare results with independent double reading. Screening mammograms of 500 cases were collected of which 125 were screen-detected cancers, 125 were interval cancers. Of al cases prior mammograms were available. All mammograms were analyzed by a CAD system which detected mass regions and assigned a level of suspiciousness to each mass. Ten experienced screening radiologists read the prior screening mammograms. Each reader read the original prior mammograms of all cases during five sessions spread out over two days. The cases were presented on conventional alternators in a random order. Mounting both current and prior mammograms is standard in the Dutch screening program. Therefore, to mimic daily screening practice as close as possible, for each case the mammogram of the screening round previous to the prior (the reference mammogram) was also mounted on the alternator to allow radiologist to judge mammographic changes over time. Independent interpretation with CAD was implemented by weighting the malignancy ratings of the radiologist with the CAD output. CAD markers on areas that were not reported by the radiologist were not used, excluding the benefit of CAD to avoid oversights. Independent double reading was implemented by averaging levels of suspiciousness assigned to each finding.

Results were evaluated using LROC analysis (figure). In total 115 cases were classified as visible priors with masses. The mean sensitivity at a false positive rate less than 10% was 39% and increased with 7.1% by CAD and with 11.3% by double reading. In conclusion, we found that the probability of CAD mass markers system can improve interpretation of mammograms, but best performance is obtained with double reading, the presence and

Researchers

• Dr. Ir. Nico Karssemeijer (Principal Investigator), Department of Radiology, UMC St. Radboud Nijmegen.
• Dr. Carla Boetes (Principal Investigator), Department of Radiology, UMC St. Radboud Nijmegen.
• Drs. Rianne Hupse, Department of Radiology, UMC St. Radboud Nijmegen.
• Maurice Samulski MSc, Department of Radiology, UMC St. Radboud Nijmegen.