Can an explainable artificial intelligence (AI) model have an impact for enhancing breast magnetic resonance imaging (MRI)?
For a retrospective study, recently published in Radiology, researchers compared an explainable fully convolutional data description (FCDD) model for anomaly detection on breast MRI to a binary cross-entropy (BCE) model across three cohorts. The cohorts included a 5,026-patient, 9,567 breast MRI exam model cross-validation development dataset; an internal test set involving 171 breast MRI examinations; and an external multicenter dataset comprised of 221 breast MRI exams, according to the study.
For the development and internal test cohorts, the researchers also examined the effectiveness of model detection tasks across balanced groups (with a 20 percent cancer prevalence) and imbalanced groups (with a 1.85 percent cancer prevalence).
In the cross-validation group, the FCDD model offered higher AUCs than the BCE model for balanced detection (84 percent vs. 81 percent) and imbalanced detection (72 percent vs. 69 percent). The researchers also noted a higher AUC for the FCDD model in external testing for balanced detection (86 percent vs. 79 percent).
“Our study revealed that anomaly detection was superior to traditional binary classification in balanced and imbalanced cancer detection tasks,” wrote study co-author Savannah C. Partridge, M.D., who is affiliated with the Fred Hutchinson Cancer Center in Seattle, and colleagues.
The study authors found that the FCDD model offered enhanced spatial agreement in contrast to the BCE model (92 percent AUC vs.81 percent AUC).
”Compared with the explanation maps of the other models, the FCDD maps demonstrated higher specificity and spatial accuracy, both desirable features for model outputs to be useful to radiologists,” added Partridge and colleagues.
Three Key Takeaways
- Anomaly detection outperforms binary classification. The FCDD anomaly detection model consistently showed superior performance over traditional binary classification (BCE) in both balanced and imbalanced breast MRI cancer detection tasks, offering higher AUCs across internal and external datasets.
- Improved spatial specificity and fewer false positives. The FCDD model provided better spatial accuracy and specificity in its explanation maps, leading to outputs more useful for radiologists and reducing false positives by an average of 25 percent.
- Higher predictive value with clinical relevance. At matched sensitivity and specificity, the FCDD model achieved double the positive predictive value of the BCE model, suggesting enhanced clinical utility for reducing unnecessary follow-up imaging.
The FCDD model also offered a higher positive predictive value (PPV) and a significant reduction in false positives in comparison to the BCE model, according to the study authors.
“At the Youden index, FCDD achieved twice the positive predictive value at similar sensitivity and specificity as the binary classification model (binary cross-entropy (BCE)) while reducing the number of false-positive predictions by an average of 25% (mean decrease of 58 of 233 false positives compared with BCE),” pointed out Partridge and colleagues.
(Editor’s note: For related content, see “Possible Real-Time Adaptive Approach to Breast MRI Suggests ‘New Era’ of AI-Directed MRI,” “Study: Abbreviated Breast MRI Offers Equivalent Accuracy to mpMRI for Women with Dense Breasts” and “Emerging AI Algorithm Shows Promise for Abbreviated Breast MRI in Multicenter Study.”)
In regard to study limitations, the authors acknowledged the emphasis on two-dimensional subtraction MIPs and cases in which only subtle abnormalities are visible on breast MRI may hamper the detection capabilities of the model evaluated in the study.
