Deep Learning Model Detects Artifacts on High-b-Value Breast DWI Slices

A convolutional neural network (CNN) detected hyperintense and hypointense artifacts on high b-value (b=1500) breast diffusion-weighted imaging (DWI) slices with AUROCs of 0.91 and 0.95, respectively, in a retrospective study of 11,806 slices. Further validation is needed.

• For multiclass artifact intensity classification, the same DenseNet121 achieved weighted AUROCs of 0.84 for hyperintense and 0.90 for hypointense artifacts.
• Radiologist evaluation of Grad-CAM heatmaps on a 5-point scale yielded mean scores of 3.52 and 3.62, indicating moderate localization capability.
• Evidence level 3; further validation is required before clinical use.

Evidence grade: 3

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.