A new radiomics model that utilizes mass amounts of data from mammography images outperformed radiologists at distinguishing malignant from benign calcifications, reported authors of a new study published in the European Journal of Radiology.
The model—which utilizes machine learning—extracted more than 8,000 features from hundreds of mammograms and classified benign calcifications from malignant ones significantly better than trained radiologists. Using the model in clinical practice, according to the researchers, could individualize patient care at a fraction of traditional costs.
“A better performance was achieved in the machine learning classifier, which means that the high-dimensional radiomic features explored more detailed information about the breast calcifications than the naked eye,” wrote Chuqian Lei, Guangdong Provincial People’s Hospital in China, and co-authors.
“Radiomics not only enables oncologists to provide highly personalized care for tumor diagnosis and identifies phenotypic subtypes, but radiomics also noninvasively provides effective decisions at a low cost.”
Mammography does a good job detecting calcifications, but its specificity for distinguishing benign from malignant findings remains low. AI-based radiomics extracts quantitative features from imaging data that humans cannot see and analyzes the relationships between those features and pathological results to arrive at a conclusion. The method has been successful in oncology, and Lei et al. hypothesized such an approach could better classify breast imaging reporting and data system (BI-RADS) category 4 calcifications.
The researchers included 212 eligible calcifications in their study, extracting 8,286 features from mammography images. The final nomogram (a chart depicting the relations between radiomic features) included six features and a patient’s menopausal state.
Results showed the new approach distinguished harmful from non-cancerous calcifications with an area under the receiver operating characteristic curve of 0.80—a “significant” difference from radiologists’ classifications, the authors noted. This was especially true for calcifications that appeared negative on ultrasound, but could be detected using mammography.
Lei and colleagues cited many strengths in their study, including a focus solely on BI-RADS 4 calcifications (considered the most difficult and suspicious category of lesions). Calcifications categorized as BI-RADS 1 to 3 and 5 have clear benign and malignant features and are “nearly meaningless to explore,” they explained.
Radiomic nomogram can identify the possible benign calcifications according to the relevant characteristics of calcifications on mammographic images. Radiomics is an emerging tool for disease diagnosis, and more studies are expected to be conducted to prove its unique value in clinical applications.