An automated system that correlates histologic and mammographic results could refine radiologists' ability to detect breast cancer, according to a multicenter study published in the February issue of the American Journal of Roentgenology.Investigators
An automated system that correlates histologic and mammographic results could refine radiologists' ability to detect breast cancer, according to a multicenter study published in the February issue of the American Journal of Roentgenology.
Investigators from Stanford University and the University of California, San Francisco used a computer model to calculate breast disease probabilities based on Breast Imaging Reporting and Data System (BI-RADS) descriptors. They built the system by identifying in the literature 25 diseases of the breast - 11 malignant and 14 benign.
The researchers tested the model by correlating histologic results from 92 consecutive image-guided breast biopsies with their corresponding mammographic findings. They spotted incorrect pathologic diagnoses with 100% sensitivity and 91% specificity and achieved a sampling error rate of 1.1%, said principal investigator Dr. Elizabeth S. Burnside.
The system's automated oversight can help radiologists identify questionable cases and reduce the incidence of sampling errors not detected while reviewing biopsy or imaging results, according to Burnside, a radiologist formerly at UCSF and now with the University of Wisconsin. The Bayesian network can easily handle and structure data in a mammographic report to assess concordance automatically, providing routine oversight to the task of imaging-histologic correlation.
Out of 43 hierarchically organized descriptors making up BI-RADS, researchers excluded five to simplify their model. They skipped skin thickening, trabecular thickening, nipple retraction, skin retraction, and asymmetric breast tissue. These findings are either late signs of breast cancer or benign features rarely observed in the population of patients undergoing percutaneous biopsy.
The study showed several limitations. Researchers did not make distinctions among 14-gauge, 11-gauge, or excisional biopsies, mostly to determine if the system could truly represent the type of population involved in clinical practice. Histologic diagnoses recorded in the system, on the other hand, are broad categorizations of the descriptors used in pathologic reports.
Study Assesses Lung CT-Based AI Models for Predicting Interstitial Lung Abnormality
September 6th 2024A machine-learning-based model demonstrated an 87 percent area under the curve and a 90 percent specificity rate for predicting interstitial lung abnormality on CT scans, according to new research.