Our objective is to develop a high-performance computer workstation incorporating image processing, pattern analysis, and computer vision techniques for enhancement, detection, analysis, and classification of mammographic features for computer-aided diagnosis of breast cancer. The workstation will be placed at the Screen Test Centre of the Alberta Program for the Early Detection of Breast Cancer in downtown Calgary for use by radiologists and health care specialists. The system is expected to improve the sensitivity, specificity, and efficiency of breast cancer screening, and possibly reduce health care costs by decreasing the need for follow-up procedures such as biopsy.
Breast cancer is a leading cause of death among women, and its incidence is rising. Recent statistics show that approximately one in ten Canadian women will develop breast cancer in their lifetime. Although curable, especially when detected at early stages, breast cancer is expected to account for 28% of incident cancer cases and 20% of cancer deaths in women.
Mammography has been shown to be effective in screening asymptomatic women to detect occult breast cancers and to reduce mortality by as much as 30% in women aged between 50 and 69 years. This apparent positive benefit has resulted in a number of leading health care societies recommending that all women aged 50-69 be screened using mammography on at least a biennial basis. In order for mass screening to be cost effective, means need to be developed to achieve it with high accuracy and speed. Even if qualified personnel are available, it is difficult for a radiologist to interpret screening mammograms in large numbers.
The debate as to the best means for analyzing large numbers of screening mammograms has led to research on the possible use of digital image processing and computer vision techniques to serve as an adjunct to visual analysis by a radiologist. Mammographic images may be digitally processed to improve their quality, to enhance features and make them more evident, to identify significant signs and quantify them, to analyze large volumes of mammograms and identify those containing significant or suspicious features, and to achieve data compression and efficient archival/ communication (teleradiology).
We have developed new image processing techniques for the enhancement of mammograms to improve the detectability of diagnostic features. One of our approaches is based on the use of adaptive neighbourhoods and region growing. This method uses regions corresponding to image features in deriving the output image. A functional contrast enhancement step is included in the method for selective contrast enhancement. Tests with difficult cases and cases of interval cancer from the Alberta Program for the Early Detection of Breast Cancer have demonstrated that the techniques can lead to improved diagnosis of breast cancer at earlier stages.
We have also developed a multi-tolerance, multi-scale, and fuzzy region growing methods to detect microcalcifications and tumors in mammograms. Further, specialized shape factors and texture features have been developed to characterize microcalcifications and tumors. The factors include moments of boundary features, Fourier descriptors, compactness, measures of concavity, spiculation index, inhomogeneity in density, as well as measures of acutance and boundary fuzzyness to characterize the diffusive nature of malignant tumors. Pattern recognition and neural network-based classification techniques have been developed to classify microcalcifications and tumors as benign or malignant. Accuracies of 85-95% have been obtained with test images.
The methods are ready for implementation on workstations for use in breast cancer screening centres. Iconographic display and data visualization techniques will be incorporated to facilitate easy comprehension by radiologists and health care technologists of the objective features computed by our techniques.
Part of a mammogram with malignant calcifications (512 x 768 pixels).
Boundaries of malignant calcifications detected by our multi-tolerance region growing method.
A 700x700 portion of a mammogram with a spiculated malignant tumor.
Result of fuzzy region growing for the spiculated malignant tumor.
Detection of breast masses: benign example (NOTE: This is an Encapsulated Postscript file of size 13 MB.) Upper left: 1000X1000 pixel segment of an original mammogram with a well-circumscribed benign mass. (Pixel size = 50 micrometers.) Upper right: Contours detected for the image. Lower left: Final contours detected by our computer method (red), superimposed with the contour drawn by the radiologist (blue). Note that one extra region has been detected by the computer method. Lower right: Analysis of the contour drawn by the radiologist- red parts indicate convex segments and green parts indicate concave segments. Benign masses typically have very few, if any, concave parts.
Detection of breast masses: malignant tumor example (NOTE: This is an Encapsulated Postscript file of size 13 MB.) Upper left: 1000X1000 pixel segment of an original mammogram with a spiculated malignant tumor. (Pixel size = 62.5 micrometers.) Upper right: Contours detected for the image. Lower left: Final contours detected by our computer method (red), superimposed with the contour drawn by the radiologist (blue). Note that six extra regions have been detected by the computer method. Lower right: Analysis of the contour drawn by the radiologist- red parts indicate convex segments and green parts indicate concave segments. Malignant tumors typically have many concave parts.
Part of a mammogram with a circumscribed, malignant tumor. The lines around the tumor boundary indicate pixels that are used to compute shape factors and measures of boundary unsharpness for tumor classification.
Part of a mammogram with a spiculated, malignant tumor. The lines around the tumor boundary indicate pixels that are used to compute shape factors and measures of boundary unsharpness for tumor classification.
The methods that we have developed are based upon characteristics of the human visual system and incorporate diagnostic indicators used by radiologists in the diagnosis of breast cancer. The techniques adapt to the wide range of details present in mammograms. These features are unique to our methods, and provide advantages in terms of improved diagnostic accuracy.
The proposed diagnostic workstation could be used in a breast cancer screening center in many possible scenarios. Foremost, it could be used as a "second reader" to assist an expert radiologist. The radiologist may or may not revise his/her initial diagnosis after obtaining the computer's diagnostic report. The Alberta Program for the Early Detection of Breast Cancer requires each case to be interpreted by at least two radiologists.
Second, the workstation could be used to pre-screen mammograms and select those that need more attention (considered to be suspicious cases). The computer techniques could be used to annotate the suspicious areas in the selected cases and draw attention of the radiologist to such areas.
Third, the computer techniques could provide quantitative analysis of diagnostic features selected by the radiologist and/or by the computer methods. The quantitative features, along with pattern classification techniques, could assist the radiologist in arriving at a more definitive diagnosis. The pattern classification techniques would incorporate details of previous diagnoses by (essentially, the expertise of) the same radiologist or an entire team of radiologists.
Fourth, the workstation could be used to train new radiologists by presenting cases with known diagnosis and offering the computer's diagnosis as a prompt or confirmation.
The results are expected to improve the accuracy of mammographic diagnosis of early breast cancer, reduce patient morbidity, and reduce health care costs.
Mammograms need to be digitized to very high spatial resolution of the order of 50x50 micrometers, resulting in data of the order of 5000x4000 12-bit pixels or 40 megabytes per image. Analysis of each clinical case requires processing of several such images. Processing large data arrays as above demands significant resources in terms of not only computing speed, but also in terms of memory and data transmission. Our techniques currently take several tens of minutes per image on Sun workstations. Practical application of the techniques in a clinical setting will demand significantly lower computing and return times.
One of our computationally-intensive contrast enhancement algorithms has been implemented on the Compaq Alpha cluster using the message passing interface (MPI). The input image was partitioned into sets of pixels of the same brightness value disregarding their location in the image. In the parallel implementation of the contrast enhancement algorithm, the master processor allots one set of pixels to each slave processor. When a slave completes processing a set, the results are returned to the master processor. The master then sends a new set of pixels to the slave for processing. This procedure continues until there are no sets of pixels left. The subdivision of the original image based on brightness values guarantees that slave processors do not process the same pixel.
The parallelism value of the problem was 16, i.e., the performance did not improve significantly when more than 16 processors were used. The performance improvement factor compared to an Ultra 1 Sun Sparc Workstation is approximately 10-20, depending on the number of processors used.
Initial tests have indicated that a mammogram may be enhanced on the Compaq Alpha cluster at MACI in a fraction of a second. Such computing speed would facilitate analysis of an entire case file with about a dozen images in the order of a couple of seconds. It thus becomes feasible to consider application of our techniques to several cases per day.
Block diagram of the proposed high-performance computing and communication system for Computer-aided Diagnosis of Breast Cancer (CADmam).
We are currently considering application of computer techniques as an adjunct to visual analysis of the mammograms by a radiologist only for difficult cases. Our long-term goal is pre-reading or screening of all cases by computer analysis and prompting or aiding the radiologist with the computer results. The goal is achievable by the computing resources at MACI. The facility may further be extended to Screen Test Centres at other cities and towns in Alberta via high-speed communication links. The proposed workstation should soon find use by radiologists and health care specialists at the Screen Test Centres of the Alberta Program for the Early Detection of Breast Cancer.