Radiomics offers improved diagnostic performance of mammography, compared to what is observed by radiologists, according to new research published in the Journal of the American College of Radiology.
“Recent studies have explored the use of radiomic features to distinguish benign and malignant lesions or molecular subtype on magnetic resonance (MR) and mammographic imaging,” wrote lead author Nan Hong, MD, PhD, of Peking University People’s Hospital in Beijing, China. “These texture features describe the intensity distributions within the lesion, capture spatial and spectral frequency patterns, and characterize the relationships among different intensity levels within the lesion.”
“A number of these features might not be visually apparent to radiologists and therefore could complement their diagnostic skill set,” Hong added.
They noted a paucity in studied that evaluate the use of radiomics for automatic differentiating between benign and malignant breast cancer in mammography. So, they sought to retrospectively evaluate if the utilization of radiomics improved the diagnostic performance of mammography, compared with that obtained by experienced radiologists.
A total of 173 patients with 74 benign and 99 malignant lesions were included in the analysis. They underwent mammography exams before chemotherapy. Radiomic features were extracted from each patient.
Four machine learning algorithms—support vector machine, logistic regression, K-nearest neighbor and Bayes classification were utilized to serve as the predictive model and an independent data set was used to validate the prediction ability of the model. The algorithm-radiomics performance was then compared to the diagnostic predictions of two radiologists.
The researchers found that 51 radiomic features remained after pre-processing. Logistic regression classification presented the best differentiation ability among the four regression models. Overall, the researchers found radiomics can adequately diagnose between benign and malignant breast tumors and can also give interpreting radiologists complimentary information that is useful in breast cancer diagnosis. Specifically, they found:
- The logistic regression model had 97.8 percent diagnostic accuracy, 97.5 percent specificity, and 98.3 percent sensitivity.
- The testing data set had 88.6 percent diagnostic accuracy, 90 percent specificity and 86.7 percent sensitivity.
- For the training data set, the two radiologists had a combined diagnostic accuracy rate of 77.2 percent, specificity of 71 percent and sensitivity of 86.2 percent.
- For the testing data set, the two radiologists had a combined diagnostic accuracy rate of 76.9 percent, specific of 69.5 percent and sensitivity of 85.8 percent.
There were some limitations. First, the cohort was a small sample size and a single-center study. Second, texture analysis was only completed on 2D images, and may be “less representative” for the entire tumor volume, compared to 3D analysis. Third, image segmentation was conducted manually, while the use of automated software is the preferred method in radiomic studies as it offers “a markedly improved efficiency and may reduce the interobserver variability in tumor delineation. And lastly, the researchers only analyzed mammography images and did not compare MRI and ultrasound imaging.
“The promising results of our study suggested that mammography images could be captured and quantified by radiomics, which offers good diagnostic ability for benign and malignant breast tumors and provides complementary information to radiologists,” the researchers concluded.