We are applying computer-aided diagnosis to static medical images (both digitally acquired and digitized from film) to recognize and characterize abnormalities, and to develop advanced analytics to approach this problem. We have applied a combination of preprocessing techniques with pattern recognition algorithms and other Machine Learning methods to aid in the detection of microcalcifications in the breast and to classify benign and malignant disease cases. A differentiating methodology in our analysis centers on the incorporation of our proprietary advanced Machine Learning technologies.

Mammography and physical examination are recognized as the current standards for early detection of breast cancer. There is conclusive evidence that routine screening mammography results in significant reduction of mortality from breast cancer for patients over 40. As a result of screening mammography, malignant lesions can be detected at an earlier stage of the disease when the prognosis is more favorable. Still, breast cancer is the second leading cause of cancer deaths in women and the second most common cancer among women, after skin cancer. It is estimated that more than 700,000 people died from breast cancer in the world in 2000, with more than 41,000 deaths in the United States. Currently, 10% to 25% of breast cancers are not detected by mammography although approximately 70% of first-pass undetected malignancies demonstrate abnormalities on prior mammograms when reviewed retrospectively. Additionally, of all cases subsequently referred for biopsy, only 20% to 30% are confirmed positive for malignancy. Microcalcifications, mass lesions, spiculation, asymmetry, and architectural distortion represent abnormalities present on mammograms that may indicate the presence of breast cancer. The use of computer-based programs to detect and analyze these abnormalities would be valuable in the diagnosis of breast cancer at an early stage when the disease remains curable. The value of computer aided diagnosis stems from the potential of specific algorithms to reduce the number of false positives, diagnosing cancer erroneously, and false negatives, or not diagnosing a present cancer, inherent in human readings of mammograms.

Currently marketed computer aided diagnostics are typically plagued by a high false positive rate because of the tendency to err on the side of being too conservative. The costs incurred in resolving such false alarms largely offset the potential benefits of automated screening. Our superior analytic tools should be able to control the number of false positives while maintaining the ability to accurately detect a very high percentage of the actual disease cases. Accomplishing this will allow the automated tools to initially be used as a second reading of the examination to validate the physician's diagnosis. Second human screenings, while cost prohibitive in general practice, have been shown to significantly improve the detection of abnormalities. An automated screening tool with near-expert screening accuracy would afford the same benefit without the costly utilization of valuable physician time. Further refinement of the automated screening tools may eventually result in primary interpretation by computer with physician time dedicated to only the "non-normal" cases.

Our preliminary results have shown great promise in the analysis of digitized mammograms. We have organized a large database into training and testing sets, achieving detection rates for microcalcifications in excess of 90%. More importantly, this has been achieved with a very low number of false positives. This is a dramatic improvement over other presently available commercial products. We plan to rapidly finalize development of this product and start marketing as an adjunct to radiologist interpretation.

We plan to incorporate additional patient data streams, including clinical data, cancer markers, and cytometry, to maximize the diagnostic capabilities for the individual patient. BIOwulf¹s drive toward automated screening capabilities should maximize clinical utility and enhance the detection of breast cancer while simultaneously reducing the number of unnecessary biopsies being performed.

Mammography will be our first application of the use of Machine Learning to interpret images. There are, however, a large number of other applications for the use of Machine Learning to increase the knowledge gained from medical images. We expect to work directly on images and image sequences already being produced by medical imaging tools, while working on methodologies to produce novel and powerful new tools for the extension of drug discovery and medical diagnostics.