The next step toward reducing the mortality of breast cancer may involve not a pathologist studying a biopsy, nor an oncologist crafting a new cancer-fighting compound, but a scientist using a computer to mine answers from digital visualizations of human tissues.
Breast cancer is the second-leading cause of death among women worldwide and, because preventing it is beyond current medical abilities, much research attention has been focused on early detection and postdiagnostic treatment. But early detection has flaws. Even mammography, the most effective tool for detecting the cancer, misses up to 30 percent of breast lesions. The missed evidence is attributed to poor-quality radiographic images and eye fatigue and oversight on the part of radiologists who read the images.
A professor who specializes in data mining in UT’s College of Business Administration is working on computer-aided diagnostic tools that show promise in increasing the ability to spot cancerous lesions in the digital images collected during mammography. Hamparsum Bozdogan, McKenzie Professor in Business, is working with the Medical Imaging Group at the University of Bologna, Italy, to develop a computer program that will catch details of radiographic images that human eyes may miss.
In the three images below, the squares indicate the location of the actual tumoral mass, whereas the circles indicate computer-aided detection. In the second and third images, the squares indicate actual tumor masses and the circular marks outside the squares indicate misidentified tumors, or false positives. Click on an image to see a larger view.
Masses are the most common lesions associated with breast cancer. In radiographic images, masses appear as thickenings of breast tissue, with sizes ranging from 3 to 30 mm. Masses also vary considerably in shape, contrast, border, and texture; and identifying a common set of features effective for every kind of mass is extremely difficult.
Bozdogan and his collaborators in Italy have developed a computer-aided diagnosis program that provides the radiologist with what he calls a “second set of eyes” to view radiographic images. Relying on statistical algorithms that seek out patterns among a variety of variables, the new data-mining techniques can detect and diagnose breast abnormalities earlier and faster than typical screening procedures.
In a series of studies of more than 1,300 clinical subjects, the techniques developed by the researchers were able to predict breast cancer malignancy 15 percent better than currently used procedures. Bozdogan is currently developing further-refined algorithms that will identify more abnormalities while minimizing the number of apparent abnormalities that are in fact healthy tissue. He hopes to improve the new procedures so that its predictive power will be 30 to 50 percent better than traditional screenings.
Decisions about diagnosis and biopsy belong with the patient and the radiologist, Bozdogan notes, but his goal is to provide a new tool that may improve diagnosis and save lives.