Researchers at the University of Oklahoma have developed an image-analysis technique that is designed to improve breast cancer detection and diagnosis.
Bin Zheng, OU electrical and computer engineering professor and Oklahoma Tobacco Settlement Endowment Trust Cancer Research Scholar, and his research team have developed image processing algorithms to generate quantitative image markers by analyzing multiple digital X-ray images and building statistical data learning-based prediction models. The goal is to develop a new quantitative image analysis method that better predicts cancer risk or cancer prognosis, which ultimately leads to help establish more effective personalized cancer screening and treatment strategies.
For example, to improve efficacy of breast cancer screening, a number of breast cancer risk factors including age, breast density, family cancer history, lifestyle and test results on some common susceptible cancer gene mutations are reviewed. Using these risk factors, several lifetime breast cancer risk assessment models have been developed and applied in epidemiology studies.
"Our study is different. We do not intend to build another lifetime risk model to compete with the existing models. We focus on developing and testing a new risk model to predict whether a woman has high risk of developing breast cancer in a near-term after a negative screening mammography," Zheng explained.
If successful, the model will help establish a new optimal personalized cancer screening model. As a result, an adaptively adjusted screening frequency and method can be applied to each woman at different time periods.
Zheng and his research team have been working to explore and identify image features and their difference, or asymmetry, between the left and right breasts. The images can be fused to build new risk models to more sensitively detect subtle image changes and/or abnormalities that are likely to lead to the development of mammography-detectable cancer in the next one to three years.
The team first identifies and computes useful image features from the two views of bilateral mammograms of the left and right breasts. Then they train statistical models (i.e., an artificial neural network) to generate a prediction score. The prediction score is the likelihood of a woman developing a "mammography-detectable" breast cancer after having a negative screening mammography examination, or classifying between malignant and benign recalls from suspicious mammograms detected by radiologists.
Routine mammography has been shown to significantly reduce mortality associated with breast cancer, but default testing a wide range of the female population brings low efficacy and other complications. Radiologists may miss or overlook a high percentage of early cancers while also generating high false-positive rates.
According to one study, in 10 years of screening more than half of women will experience a false-positive recall, and 9 percent will receive a false-positive biopsy. While that may be relieving to the patient considering the alternative, false-positives are not harmless. Misdiagnosis can cause psychosocial issues, increased radiation, pain related to a biopsy and increased health care costs.
"Our preliminary study results demonstrate that our new near-term risk prediction model based on a computer-aided detection scheme of four-view mammograms yielded a substantially higher discriminatory power than other existing known risk factors to predict near-term cancer risk," Zheng said.
The advanced prediction could help the medical community improve cancer screening efforts by focusing on women at greatest risk for developing breast cancer in the near-term and also reducing the number of women harmed from false-positive results.
"The ultimate goal is to develop a personalized cancer screening," Zheng explained. "Since cancer development is a progressive process, our new model focuses on detecting this dynamic process from the images and then improving the near-term breast cancer risk stratification among the women who participate in mammography-based breast cancer screening."
As a result, only the small percentage of women stratified into the group of high risk in near-term should be more frequently screened, while the vast majority of women stratified at average or lower near-term cancer development risk could be screened at longer intervals - for example, every two to five years. This would increase cancer detection rate by focusing radiologists' attention more on a small fraction of high-risk women by reducing the missed and/or overlooked subtle cancers, while also reducing the annual screening population and associated false-positive recalls among the vast majority of women with low near-term cancer risk.
Zheng's research is currently supported by research grants from the National Cancer Institute and the Stephenson Cancer Center at the OU Health Sciences Center. His research team actively collaborates with other clinical and bioengineering researchers from the OU College of Engineering, OU Health Sciences Center, University of Pittsburgh and Mercy Women's Center in Oklahoma City to develop and test new quantitative image analysis methods to improve efficacy of breast cancer screening using mammography and breast magnetic resonance imaging, prognosis prediction of early-stage lung cancer and clinical benefit assessment of clinical trials for testing new chemotherapy drugs for treating ovarian cancer.
"We try to bridge the research and application gap between using genotype biomarkers and phenotype image markers," Zheng explained. "Our study results demonstrated that quantitative image markers could provide useful and supplementary information to the existing biomarkers and/or risk prediction models. Fusion of these two types of markers has potential to yield significantly higher performance in predicting cancer risk and prognosis."
University of Oklahoma College of Engineering