In external validation testing, the deep learning and combined models provided over 40 percent higher sensitivity rates (94 and 96 percent respectively) in contrast to 53 percent for the clinical model, according to the study authors. They also pointed out the deep learning and combined models had 11 percent higher specificity (92 percent vs. 81 percent) and approximately 30 percent higher accuracy (93 and 95 percent respectively vs. 63 percent) in comparison to the clinical model.
“These findings underscore the potential of deep learning to minimize overtreatment and ultimately improve patient outcomes,” wrote lead study author Wenjie Xu, M.D., who is affiliated with the Department of Radiology at the Tangde Hospital of Zhejiang Province in Zhejiang, China, and colleagues.
While noting that associated mass, asymmetry and architectural distortion accompanying microcalcifications often suggests invasive breast cancer, the study authors pointed out these features were present in 27 of the DCIS cases in the cohort.
Acknowledging that radiomic applications have made inroads for differentiating benign and malignant breast lesions, the researchers said the dependence of radiomics on low-level phenotypic features and labor-intensive manual delineation hamper their utility.
Three Key Takeaways
- Deep learning models markedly outperformed traditional clinical models. in differentiating DCIS and IDC from suspicious microcalcifications, achieving higher AUC (97 percent vs. 67 percent), sensitivity (94–96 percent vs. 53 percent), specificity (92 percent vs. 81 percent), and accuracy (93–95 percent vs. 63 percent).
- Improved diagnostic performance may help reduce overtreatment. The study findings show more accurate preoperative distinction between DCIS and invasive disease.
- Deep learning addresses limitations of radiomics and clinical assessment. The researchers emphasized that deep learning offers automated, high-level feature extraction that enhances classification and prediction accuracy.
However, deep learning offers significant potential for bolstering preoperative diagnosis in cases involving challenging microcalcifications, according to the study authors.
“By contrast, deep learning leverages models with numerous hidden layers and extensive training data to extract valuable features, enhancing classification and prediction accuracy,” maintained Xu and colleagues.
(Editor’s note: For related content, see “Large Mammography Study Affirms Value of AI in Breast Cancer Detection,” “Predicting DCIS Upgrade to Invasive Breast Cancer: Can Contrast-Enhanced Ultrasound Have an Impact?” and “Reducing Mammography Workload by Nearly 40 Percent? What a New Hybrid AI Study Reveals.”)
In regard to study limitations, the authors acknowledged the relatively small sample size (75 patients) for the internal and external validation cohorts. The researchers also noted use of the same imaging equipment at both centers may thwart extrapolation of the results to other facilities with different scanners. The study authors conceded potential patient selection bias with the exclusion of women who did not have surgery for biopsy-diagnosed DCIS.
