From Mammograms

Yulei Jiang 1 Techniques for Classifying Breast Lesions 342

The University of 1.1 Classification Based on Human Perceptual Features • 1.2 Classification Based on Computer

Chicago Extracted Features • 1.3 Classifiers

2 Performance of Computer Classification 349

2.1 Classification Performance Based on Human Perceptual Features • 2.2 Classification Performance Based on Computer-Extracted Features • 2.3 The Potential of Computer Classification to Improve Mammogram Interpretation

3 Effect of Computer Classification on Radiologists' Diagnostic Performance 350

3.1 Diagnostic Aid Based on Human Perceptual Features • 3.2 Diagnostic Aid Based on Computer-Extracted Features

4 Summary 353

References 354

Breast cancer remains the most frequently diagnosed and the second most lethal cancer for U.S. women. In 1999, an estimated 175,000 women were expected to be diagnosed with invasive breast cancer (not counting 39,900 new cases of ductal carcinoma in situ) and 43,300 were expected to die from it. The odds of developing invasive breast cancer over one's lifetime are 1 in 8 [1]. Until an effective prevention mechanism becomes widely available, early detection followed by effective treatment is the only recourse for reducing breast cancer mortality. The effectiveness of breast cancer treatment depends strongly on cancer progression. Compared to advanced cancers, patients diagnosed with minimal cancers have significantly better 20-year survival rate [2-4]. Currently, state-of-the-art mammo-graphy is the most effective imaging modality in breast cancer detection, particularly in finding nonpalpable small cancers (in situ or invasive cancer less than 1 cm in diameter) [5-9]. Large-scale, randomized, controlled clinical trials have shown that annual screening of asymptomatic women reduces breast cancer mortality by 20-30% [6,10,11]. The results of these studies have led to the recommendation and systematic implementation of screening for breast cancer in the United States [12,13], Canada, and many European countries.

The important role that mammography plays in breast cancer detection can be attributed largely to technical improvements [14] and the dedication of radiologists (mam-mographers) to breast imaging. However, despite remarkable advances, current state-of-the-art mammography still faces challenges. It is known that mammography is unable to detect all breast cancers and is unable to reliably differentiate between malignant tumors and benign lesions [15-21]. The effectiveness of breast cancer screening is compromised by not being able to detect all breast cancer cases. Inability to reliably differentiate benign lesions from malignant tumors not only subjects healthy women to unnecessary trauma and anxiety, it also increases the cost of breast screening [22,23]. Estimates from analysis of large series of mammograms show that the sensitivity of mammography (the fraction of breast cancers that are detected) is as high as 85-95%, but positive predictive value at biopsy (the fraction of lesions biopsied that are proven to be malignant) only ranges from 15 to 30% [15].

Causes that limit the sensitivity of mammography include poor differentiation in radiographic appearance between cancerous tissue and normal breast parenchyma [24], inappropriate imaging techniques such as failure to include the breast lesion in the mammogram, and observational factors such as fatigue, distraction, and the satisfaction-of-search effect, i.e., stopping the systematic search for lesions prematurely after an abnormality is found. Current efforts focusing

Copyright © 2000 by Academic Press.

All rights of reproduction in any form reserved on improving the sensitivity of mammography include the development of full-field digital mammograms (FFDMs), which better utilizes available radiation contrast between tumor and normal breast parenchyma [24,25]; better quality control and the accreditation of mammography clinics; and double reading [26-28], as well as standardization of the interpretation and reporting of mammograms [29]. Double reading, in which a mammogram is read by two radiologists, has been shown to improve sensitivity by as much as 15%: Although each radiologist may miss a small number of cancers, those missed might be detected by the other radiologist. However, because each mammogram is read by two radiologists, this approach is inefficient from a practical point of view because radiologists' productivity is reduced. As an alternative to double reading by two radiologists, computer-aided diagnosis, in which a computer acts as the "second reader,'' has been proposed [30-32]. Computer-aided diagnosis (CAD) is defined as a diagnosis made by a radiologist who considers a computerized analysis of the mammogram in his or her diagnostic decision-making. Computer-aided detection can help improve sensitivity when the computer identifies lesions missed by a radiologist. This potential of CAD has been shown in several laboratory tests for the detection of breast lesion [33,34], lung nodule [35], and pulmonary interstitial infiltrates [36].

During the past 10 years, extensive research has been devoted to developing computer-aided diagnostic schemes for the detection of various radiographic abnormalities [31,34,3741], which are reviewed in [30,42,43]. Reported performance of computer techniques has steadily improved over the years, with some reaching a level thought to be sufficient for clinical application. The performances of automated techniques are often characterized by the numbers of true lesions (sensitivity) and false positives detected by the computer, in a set of radiographs known as a database. For example, Nishikawa and colleagues reported 87% sensitivity and approximately 0.5 false-positive detection per image for clustered microcalcifications from a database of 78 mammograms, half containing subtle clustered microcalcifications [40]. Giger et al. reported 92% sensitivity and 2 false-positive detection per image for breast masses, from a database of 154 pairs of bilateral mammograms, 90 pairs containing masses [41]. As automated detection methods begin to mature, initial clinical evaluation of these techniques has begun [44].

Increasing the detection rate of abnormalities solves only part of the problem; mammography still faces the additional challenge of improving its ability to differentiate malignant and benign breast lesions. Figures 1 and 2 show examples of malignant and benign breast lesions in mammograms. Low positive predictive value is caused by the morphological similarities between breast cancer and many benign diseases, by observational factors such as inexperience, and by medical-legal constraints. Imaging techniques which are not based on x-rays, such as ultrasonography (US), magnetic-resonance imaging (MRI) [45,46], and recently radionuclide imaging [47, 48] are being investigated to explore different avenues for better visualization of breast cancer.

Although these new techniques focus on better characterization of breast cancers, variability in observer performance needs to be addressed [49-52]. Computer-aided diagnosis can potentially play an important role in this regard. CAD may be useful in improving sensitivity as well as positive predictive value of mammography-directed biopsies (also known as positive biopsy yield). Sensitivity can be improved when a radiologist changes his or her recommendation regarding malignant tumors from follow-up to biopsy as a result of a computer's positive assessment. Positive biopsy yield can be improved when a radiologist changes his or her recommendation regarding benign lesions from biopsy to follow-up as a result of a computer's negative assessment, while retaining the correct recommendations of biopsy for all malignant tumors.

The purpose of this chapter is to provide an overview of the motivation, approach, performance, and potential clinical impact of computer classification of malignant and benign breast lesions. Numerous investigators contributed to this field, which will be illustrated in this chapter by describing the work and findings of some [53-59], while the work of many others can be found in references [60-65]. Good reviews of the earliest work in the 1970s and 1980s are also available [42,43].

The rest of this chapter is organized in three sections. First, we describe the techniques of computer classification of breast lesions. Second, we summarize the state-of-the-art performance in computer classification. Third, we describe in more detail two observer studies that provided evidence that computer classification can potentially alter radiologists' diagnosis and improve their accuracy. Throughout the chapter, we will illustrate computer classification by examples. Whenever appropriate, we will critically review the advantages and disadvantages of competing approaches. We will also occasionally identify common pitfalls in this active field of research. To help organize the three sections, it will be necessary to break up the description of research work used as examples into parts, and we shall return promptly to the same examples throughout the chapter.

Was this article helpful?

0 0

Post a comment