Conclusions

CADiagnosis is an area that merges the fields of signal processing and pattern recognition for the creation of tools that can have a significant impact in health care delivery and patient management. CADiagnosis algorithms usually involve several modules that need to be separately optimized and validated for an overall optimum performance. In this chapter, we presented a CADiagnosis methodology for the differentiation between benign and malignant breast calcification clusters in mammograms. We specifically looked into one of the aspects of the algorithm, namely the impact of segmentation in the overall classification process, and the role of multiresolution analysis in the segmentation process.

Our classification approach was based primarily on morphological and distributional features of mammographic calcifications and, hence, the role of segmentation was particularly important in the overall implementation and performance. Knowing the limitations of image segmentation techniques that were further exaggerated by our additional challenge to preserve morphology and distribution, we developed two multiresolution filters that were able to yield successful and clinically promising results. Although far from perfect segmentations, the symmlet wavelet and the donut filter adequately preserved the characteristics of the calcifications as required by the overall algorithm's design. A new filter, labeled as the "donut filter," was introduced for mammogramprocess-ing that seems to offer a robust solution to the problems associated with the detection and segmentation of mammographic images. The new filter was not utilized to its full potential and several implementation pathways remain to be explored. Its initial testing, however, yielded promising results and its usefulness could go beyond mammography to other medical imaging applications.

An important question at the end of the experiments presented here is whether similar classification performance can be achieved, either with the symmlet wavelet or the donut filter, for images generated from various sources. For example, for images generated by different film digitizers (laser-based vs. charge-couple-device-based systems), or by different imaging systems (screen/film vs. direct digital systems), or with different resolution characteristics (pixel size and bit depth). Preliminary work with different data types suggests that similar classification results may be obtained if a standardization process is applied to the images prior to processing. As long as pixel size and depth are within acceptable ranges for CADiagnosis applications in mammography, a standardization algorithm can easily convert the characteristics of any set of data to those for which the CADiagnosis system was initially trained and optimized keeping performance consistent [20, 73].

An interesting spin-off application of our initial development originated from the FP impact observation on classification performance. We found that classification results could be used as an indirect measure of segmentation quality particularly when the classification scheme is based solely on morphological and distributional characteristics like the one described here. Segmentation evaluation is one of the most challenging issues in medical image processing. It usually requires objective and accurate "ground truth" or "gold standard" information that is often unattainable in medical imaging where the human observer is commonly the only source of "ground truth" information. Using the classifier's output for indirect segmentation validation may offer an advantage over more traditional techniques that use absolute measures of shape and size and require exact ground truth information. After all, it is the clinical outcome that is important in these applications.

Finally, the described CADiagnosis scheme seems to be amendable to a variety of applications beyond breast cancer screening and early diagnosis. The input feature set and classification output could be modified and expanded to address problems associated with the diagnostic patient and specific breast disease types involving calcifications, e.g., ductal carcinoma in-situ, for the development of computer tools that go beyond detection and diagnosis into the domains of prognosis, patient management, and follow-up.

0 0

Post a comment