 |
 |
|
|
||
PLASTIC SURGERY NEWS, SURGERY NEWS, COSMETIC SURGERY NEWS Current NEWS  |
Computer-Aided Detection in Diagnostic Mammography
Detection of Clinically Unsuspected Cancers
Sherry A. Butler; Richard J. Gabbay; Deborah A. Kass; Daniel E. Siedler; Kathryn F. O'Shaughnessy; Ronald A. Castellino |
Abstract and Introduction
Abstract
Objective: We had two objectives: to determine the percentage of women presenting with clinical findings whose diagnostic mammogram led to detection of a breast cancer at a site distant from the original clinical complaint and to assess the performance of computer-aided detection (CAD) on diagnostic mammography.
Materials and Methods: Three institutions contributed consecutive cases in which a mammogram was obtained to evaluate a clinical finding, after which a histologic diagnosis of breast cancer was made. Clinical data and the mammograms were reviewed to determine the nature of the clinical findings and to document the location and characteristics of 212 biopsy-proven cancers in 197 patients who met the study criteria. Standard four-view breast mammograms were then analyzed by a CAD system.
Results: The most common clinical finding was a palpable mass (90%, 177/197), with nipple discharge (5%, 9/197), focal tenderness or pain (2%, 5/197), and miscellaneous complaints (3%, 6/197) also noted. Two separate cancers were found in 7.6% (15/197) of the cases. In another 7.6% (15/197) of the cases, the single diagnosed cancer was not at the location of the specific clinical finding. The CAD system correctly marked 87% (26/30) of those cancers that were clinically unsuspected (i.e., not at the location of the clinical finding).
Conclusion: Breast cancers occurred at locations other than the site of the presenting clinical finding in 15% (30/197) of patients undergoing diagnostic mammography in whom a cancer was detected. CAD identified 87% of these incidentally detected cancers and may therefore be useful as a detection aid to the radiologist when interpreting diagnostic mammograms.
Introduction
Interest is growing in the use of computers as aids to radiologists in the perception of findings that may represent specific disease processes. This application has been termed computer-aided detection (CAD). This technology uses software to search an image and to provide prompts (usually in the form of marks or symbols on a cathode-ray tube, flat-screen panel, or paper printout) to alert the radiologist to features that might otherwise have been overlooked. With current CAD technology, the radiologist must then evaluate these prompted areas to determine whether the findings are sufficiently compelling to warrant further investigation.
Screening mammography has been the major focus of CAD development because the well-reconized rate of false-negative interpretations of these mammograms is approximately 20%, as determined by prospective clinical experience and test studies.[1,2] Current CAD algorithms are reported to mark over 98% of microcalcifications and 86% of other lesions (masses, including architectural distortions) due to biopsy-proven breast cancer.[3,4] With such technology, up to 20% more breast cancers may be detected on screening mammography.[5–8]
For these reasons, CAD has been increasingly used as an aid in screening mammography. However, the utility of CAD for evaluating diagnostic mammograms is unknown. (For our study, we defined diagnostic mammograms as standard four-view mammograms obtained in patients who presented with clinical signs or symptoms that aroused suspicion of breast cancer.) Currently, CAD technology serves to direct the attention of the radiologist to features that might inadvertently be overlooked. Clearly, CAD is not needed to evaluate the specific anatomic site in the breast producing the sign or symptom because this location is already known to the radiologist. However, the remainder of the breast tissue of the ipsilateral breast and the entire breast tissue in the contralateral breast are actually being screened during a diagnostic study, in that these areas are asymptomatic. Furthermore, the appropriate focus of the radiologist in evaluating the specific site of clinical concern might lead to additional inadvertent oversights based on the well-recognized "satisfaction of search" phenomenon.[9]
Between 1.0% and 2.6% of patients with newly diagnosed breast cancer also have a second (synchronous or multicentric) breast cancer, in either the same or contralateral breast.[10] Also, although most patients who present with a specific clinical sign or symptom do not have breast cancer, at times a cancer is detected at a site other than the area of initial clinical concern. For these reasons, we undertook a retrospective study to determine the percentage of women presenting with clinical findings whose diagnostic mammograms led to the detection of a biopsy-proven breast cancer and who either had a cancer at a site unrelated to or distant from the location of the specific clinical finding or had two biopsy-proven cancers. In addition, we evaluated the ability of a CAD system to correctly mark the locations of these clinically unsuspected breast cancers.
Materials and Methods
Three institutions retrospectively identified all cases within a defined period (Kaiser Permanente facilities in California: San Francisco, February 1994–August 1999; Redwood City, July 1996–November 1998; and Santa Clara, October 1996–December 1997) in which a diagnostic mammogram was obtained to evaluate a clinical finding (symptom or sign) and a biopsy-proven breast cancer was eventually diagnosed. Institutional review board approval was obtained for this retrospective study with a waiver for informed consent because patient-identifying data were removed from all material. As discussed earlier, we defined a diagnostic mammogram as a standard four-view mammogram obtained in patients who presented with clinical signs or symptoms that aroused suspicion of breast cancer. Any subsequent tailored views of areas of concern, such as compression spot views, were not included in the analysis because CAD algorithms were not developed to analyze tailored views. The clinical finding leading to the performance of diagnostic mammography was tabulated. All cases of breast cancer were histologically documented (cases diagnosed as only lobular carcinoma in situ at pathologic examination were not included).
Cases meeting these criteria were excluded if the biopsy-proven cancer was not mammographically evident (as determined by the study radiologist at the site) or if the mammograms were not available for review. Patient questionnaires, mammograms, and radiology and pathology reports were reviewed to determine the nature of the clinical findings and to document the location and characteristics of all biopsy-proven cancers that met the study criteria. The location of the clinical finding was, in most cases, indicated by a radiopaque marker at the site of a palpable finding.
In this study, we defined a clinically unsuspected cancer as either a cancer that was first discovered on the four-view diagnostic mammogram and was at a site unrelated to that of the clinical sign or symptom that prompted performance of diagnostic mammography or a second cancer if the first cancer was discovered because of the clinical finding. The study radiologist categorized a case as having two cancers if the cancers occurred in different breasts, if they occurred in the same breast but in different quadrants, or if they had different histologies. The relative locations of the clinical findings and the cancers were measured, using the mammographic view in which the locations were the most distant.
The diagnostic mammograms were retrospectively analyzed by a CAD system (ImageChecker M1000 V2.2, R2 Technology). Each case was first digitized and then analyzed by proprietary algorithms that search the images for suspicious areas that might indicate masses, which include architectural distortions (marked by an asterisk) or microcalcifications (marked by a triangle). The location of the marks displayed by the CAD algorithm was compared directly with the location of the documented biopsy-proven cancer, as outlined by the site radiologists on the original mammograms before CAD analysis. The algorithms were judged to have correctly marked the lesion if the center of the CAD prompt was located within the outlined cancer site and with the correct characterization (i.e., microcalcification or mass) marker on at least one mammographic view.
Results
A total of 197 patients met the study criteria (Fig. 1). Fifteen patients had two separate biopsy-proven cancers, resulting in a total of 212 (197 + 15) biopsy-proven cancers. Invasive cancer was present in 92% (182/197) of the patients. For this study, four mammographic views were available for review in 183 cases, three views were available for one case, and two views were available in 13 cases, leading to a total of 761 images reviewed for the study.
The most common clinical finding prompting performance of the diagnostic mammography was a palpable mass (90%, 177/197), followed by nipple discharge (5%, 9/197), and focal tenderness or pain (2%, 5/197). The remaining 3% (six cases) had nonspecific findings of tenderness or pain that were not localized to a particular region of a breast.
A single biopsy-proven cancer was identified in 92.4% (182/197) of patients. In 7.6% (15/197) of the patients, two separate cancers were diagnosed—one at the location of the clinical finding and a second cancer at a separate site (Table 1). The mammographic appearance of nine of these second nonpalpable cancers was that of a mass or architectural distortion, whereas the mammographic finding in the remaining six cancers was microcalcification clusters.
One hundred eighty two patients had one cancer, 15 of whom (8.2% [15/182], or 7.6% [15/197] of the total number of patients in the study) had a cancer that was not at the location of or was unrelated to the clinical finding. For this study, cancers detected in patients with nonspecific findings (e.g., bilateral breast pain) were considered to be unrelated to the clinical finding (Table 2). The mammographic appearance of eight of these cancers was that of a mass or architectural distortion; the remaining seven cancers were characterized as microcalcification clusters. Thus, 30 clinically unsuspected cancers were diagnosed at a site unrelated to the presenting clinical finding. It is this group of 30 cancers that allowed a measurement of the potential benefit of CAD in the diagnostic mammography setting.
The CAD system marked 87% (26/30) of the biopsy-proven cancers that were not at the location of or were unrelated to the specific clinical finding that had prompted the performance of diagnostic mammography. The CAD system marked 88% (15/17) of the masses and 85% (11/13) of the microcalcification clusters in these 30 clinically unsuspected cancers (Figs. 2A, 2B, 2C, 2D and 3A, 3B, 3C, 3D). The average number of false marks per four-view mammogram (for all 197 cases) was 2.7.
Discussion
Prior studies reporting on the performance of CAD systems have specifically focused on its application to screening mammograms. In part, this emphasis was an attempt to address the known false-negative rate of approximately 20% caused by observational oversights (as compared with false-negative mammograms due to mammographically occult tumors). In retrospective studies, the sensitivity of CAD systems in correctly marking cancers that had been detected by the radiologist interpreting the screening mammogram is approximately 90% (98% for microcalcifications and 86% for masses).[3] In prospective clinical studies, the reported increase in cancer detection using CAD has ranged from approximately 20% in community practice settings[5,6] to 6.6–8.0% in academic medical centers.[7,8]
Our study differs from others that have reported on the ability of CAD systems to mark breast cancers in that the images that we analyzed were exclusively diagnostic mammograms. In this clinical setting, the radiologist is already alerted to the fact that the woman has a clinical finding (symptom or sign) in a specific area of the breast. In our opinion, the use of CAD to direct the radiologist's attention to a specific site is of dubious added value because the radiologist's attention is already drawn to that site by a radiopaque marker or other clinical information depicting the location of the palpable mass or focal pain, for example. However, for those cases in which the clinical finding is not well localized, a CAD marker at a specific location may be of benefit.
It is important to note the remainder of the breast tissue in the ipsilateral breast and in the asymptomatic contralateral breast still needs to be screened for a possible cancer. Although radiologists are trained to avoid detection errors (e. g., satisfaction of search), the use of CAD systems to help direct the attention of the radiologist to all suspicious areas in the image could be a potential benefit. We therefore focused in this study on the performance of CAD to detect cancers unrelated to the location of the clinical finding that prompted performance of diagnostic mammography.
The 30 clinically unsuspected cancers that were detected at a site unrelated to the presenting clinical finding allowed us to measure the potential benefit of CAD in a diagnostic mammography setting. The CAD system successfully identified 87% (26/30) of these radiologist-detected biopsy-proven cancers. This CAD detection sensitivity is similar to that reported for a large series of consecutive breast cancers that were detected by radiologists in their screening programs (90%).[3] The false marker rate of 2.7 marks per four-view diagnostic case in our study, however, is higher than the 2.0 false marks per normal screening case reported elsewhere.[3] This result is to be expected because case sets consisting of known cancer cases have more suspicious areas marked (R2 Technology, unpublished data) than case sets of clinically confirmed normal examinations, possibly because the images in the cancer case set are more likely to be complex. Also, we should note that the version of software used in this study is not the latest available. CAD software is continually being improved, and later versions might perform better or worse in this patient population.
An important limitation of this study is the retrospective nature of the analysis, in that all breast cancers were radiologist-detected. In addition, we have no proof that a radiologist prospectively using the CAD system would take action on the basis of one of the CAD marks. Thus, the potential benefit of CAD in a diagnostic setting requires verification in prospective studies, just as the benefit of CAD in a screening mammography program has been established with prospective studies[5–8] after initial retrospective studies.
Another limitation of the study is the potentially inexact clinical and mammographic correlation of findings because we retrospectively analyzed patient records. Not all palpable clinical findings were noted with radiopaque markers. To be conservative, we assumed that the cancer was at the site of the clinical finding unless there was clear evidence otherwise—such as the cancer being detected in the opposite breast from the one noted in the clinical finding or the cancer being very subtle and unlikely to be palpable. For cases identified with two documented cancers in the same breast and with the same histology, it is possible that the two separate mammographic findings were identifying different portions of the same cancer.
In summary, 15% (30/197) of diagnostic mammograms in which a cancer was diagnosed had a cancer detected at a location distant from or unrelated to the site of the presenting clinical finding. CAD marked 87% of these cancers, suggesting that CAD could potentially be used to help the radiologist screen the remainder of the breast tissue on diagnostic mammograms
Address correspondence to K. F. O'Shaughnessy, R2 Technology, 1195 W Fremont Ave., Sunnyvale, CA 94087.
http://www.medscape.com/viewarticle/491671
