Data from a recent study of biopsy-proven cases show that an interactive CAD system can increase accuracy in the characterization of malignant and benign breast lesions and can reduce the number of negative biopsies.

Worldwide, breast cancer is the most frequent form of cancer in women [1]. Together with self-examination, screening can identify possible breast abnormalities, with ultrasound imaging being used as a diagnostic tool to characterize such abnormalities or lesions. Recommendations for further imaging or biopsy are thus made based on ultrasound imaging. In the United States, the American College of Radiology has developed, and recommends, that the Breast Imaging Reporting and Data System (BI-RADS) be used as the lexicon to evaluate breast abnormalities [2]. Several other organizations worldwide also offer standard practices for the characterization of breast lesions. Prior to BI-RADS, many criteria were used globally; of these the Stavros Criteria describe lesions in terms of malignant and benign descriptors as shown in Table 1 and 2. The 1995 Stavros paper shows that each individual lesion characteristic contributes a certain degree of accuracy [3]. The overall goal is to have a diagnostic accuracy greater than 98% for both malignant and benign breast lesions; this was achieved by Stavros, whose paper also shows that individual lesion descriptors can contribute from 24% to 80%. The Stavros paper shows that there is no one lesion description that can achieve the accuracy goal. However, using all of the lesions descriptions in concert with each other can achieve a high diagnostic accuracy and meet the goal for diagnostic accuracy.
There can be variability among readers when using lesion classification systems [4, 5]. Variance can result from not using all of the lesion descriptions as well as from incorrectly defined classifying terms; such variance can lead to lowered diagnostic accuracy which can result in an increased number of negative biopsies [4]. If lesion classification could be effectively automated through use of CAD, then perhaps the variability would be reduced, thus changing outcomes by reducing the number of negative biopsies.

Computer-aided-detection

CAD aids physicians by including more lesion description terms and by standardizing the way in which the terms are used. In breast ultrasound, CAD finds the edge of the lesion and then analyzes it by using information from both inside and outside the lesion matrix in order to determine the terms which can be used to describe the lesion. CAD uses a computer algorithm based upon expert reader’s identification, characterization of lesion characteristics and BI-RADS scores. Medipattern’s B-CAD algorithm is based on over 15,000 lesion descriptors read by four experts. The resultant algorithm uses the expert domain knowledge to establish mathematical definitions of the BI-RADS features.
Using CAD to characterize a lesion is a new feature of CAD technology, which up till now was most often associated with CAD marks that show where there may be an area of concern. B-CAD is an interactive CAD system and works with the operator to characterize the lesion after the operator has determined that it is questionable. The operator simply clicks on the area of concern and the CAD system then outlines it, describes it and measures it. CAD finds the edge of the lesion which is referred to as segmentation. The algorithm actually finds several candidates for the edge of the lesion, using the best fitting outline of the lesion boundary (goodness of fit) to determine the most likely candidate for segmentation. It then characterizes the lesion using the BI-RADS lexicon. BI-RADS features are determined through separate algorithms for each feature which look at the lesion edge, matrix and the region surrounding the lesion as defined for each feature. The algorithm then uses the mathematical definition for each characteristic and selects the most likely term to describe the lesion. CAD further generates a BI-RADS score which suggests either that the lesion is probably benign or highly suggestive of malignancy based upon the features used to describe each lesion.
The CAD results are included in a report or worksheet format which standardizes the content and the language used in the report to improve consistency and ultimately communication. A sample of a lesion with and without CAD is shown in Figure 1. Medipattern’s B-CAD, as implemented in the workflow solution from the German company TomTec Imaging Systems was used to generate the CAD results.
The example in Figure 1 is biopsy-proven colloid cancer. In addition to showing the lesion with and without CAD, the system also fills in descriptive terminology to give a complete synopsis of the lesion and then provides the information is a standardized report.
One of the true tests of any type of CAD is to look at the results generated for benign pathology, since solid nodules can be benign or malignant and it is often difficult to determine the difference between the two. Figure 2 shows CAD applied to a biopsy-proven benign lesion. In this case the irregular part on the edge of the lesion could be an indication of a malignant lesion. The CAD is not fooled and even though it segments the irregular part of the lesion correctly, the CAD continues to correctly characterize the lesion. This characterization step is critical to being able to obtain a correct interpretation. The CAD algorithm contains definitions of just how round a lesion has to be to be considered round, how irregular to be considered irregular and how oval to be considered oval. The CAD carries these distinctions in definitions throughout the rest of the lesion descriptors. One of the benefits of CAD is the assistance that it gives in distinguishing between the choices in lesion descriptors.

Extended study of biopsy-proven lesions

The cases shown in Figures 1 & 2 are just a few examples; the real test is of course to see how well the concept performs in a blinded study of many lesions. A group from the Joint Department of Medical Imaging, University of Toronto, Canada recently carried out a study designed for this purpose. The results were presented at RSNA 2010 and showed clearly that interactive CAD could achieve high sensitivity and could more accurately determine if a lesion needed to
be biopsied [6].
The study covered a wide range of cancer types and benign nodules from a diverse population and involved retrospective Research Ethics Board-approved review of breast biopsy database. The study blinded the reader to the original outcomes and subsequent pathology results. and examined 320 consecutive lesions which had had ultrasound (US) guided breast biopsy in 2006. Of these lesions, 54 were excluded from evaluation: 3 patients had no images associated with the procedure; 12 patients had no pre- or post-biopsy image to evaluate (despite an image demonstrating a biopsy needle in situ); 37 patients had the ultrasound image linked to another mammographic study and so could not be loaded into the B-CAD "study list" and 2 patients were deemed by the CAD as below the image quality level required for successful segmentation. The remaining 266 histologically proven lesions comprised 164 breast carcinomas and 102 benign breast lesions from patients aged 16 to 92 years.
The paper thus describes the performance of interactive CAD for biopsy-proven malignant (n=164) and benign (n=102) breast disease. The CAD system showed a sensitivity of 98%, specificity of 93%, PPV of 96% and NPV of 98%. The use of interactive CAD analysis of the 102 benign breast lesions using interactive CAD could have resulted in 95 fewer biopsies being carried out. There were 2 cases of malignancy (both invasive ductal carcinomas) which were segmented as BIRADS 2, 3 or potentially benign by the interactive CAD. As with all CAD systems, the output information is a suggestion; the physician must interpret each lesion on its own merit.
All of the cases in the study had originally been interpreted in practice and recommended for biopsy. The original physicians had read these cases and determined that all were BI-RADS category 4a or higher. Of the 266 cases, 102 proved to be negative or benign and 164 proved to be malignant by pathology — a positive biopsy rate of malignancy in 62% without CAD. If CAD had been used, the study shows that it would have recommended a biopsy for 169 lesions with 162 being found correctly. This gives a positive biopsy rate of 96% — an increase of over 30% — and spares patients from biopsies that would have proven negative (and also lowers the burden on the healthcare system), while maintaining a higher standard of excellence in practice.
As a result of using the interactive CAD, the readers increased their accuracy in the characterization of benign lesions resulting in a decrease of the negative biopsy rate while maintaining high accuracy for detection of malignant lesions. The data suggest that Medipattern’s interactive B-CAD algorithm as implemented by TomTec Imaging Systems is a promising addition to current clinical breast sonographic imaging.

References

1. Tyczynski J et al. Breast Cancer in Europe. ENCR Fact Sheets Dec 2002, Vol 2. (www.encr.com.fr)

2. Breast Imaging Reporting and Data System, BI-RADS, Lexicon. American College of Radiology, 2004 edition.

3. Stavros AT et al. Solid Breast Nodules: Use of Sonography to distinguish between Benign and Malignant Lesions. Radiology 1995; 196:123- 134.

4. Baker et al. Sonography of solid breast lesions: observer variability of lesion description and assessment. AJR Am J Roentgenol 1999; 172: 1621–1625.

5. Lazarus E et al. BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value Radiology. 2006 May; 239(2): 385-91. Epub 2006 Mar 28.

6. O’Donoghue PM et al. Retrospective evaluation of computer-aided-detection (CAD) system in characterization of sonographically depicted breast lesions RSNA 2010