Truace
Subscribe

Slice-Wise Quality Assessment of High b Value Breast DWI via Deep Learning-Based Artifact Detection

Diffusion-weighted imaging (DWI) can support lesion detection and characterization in breast MRI. Therefore, DWI is increasingly being incorporated into breast MRI protocols to address some of the shortcomings of routine clinical breast MRI. However, especially high b value DWI can be prone to intensity artifacts that can affect diagnostic image assessment. To detect hyper- and hypointense artifacts on high b value DWI (b = 1500 s/mm2) using deep learning, employing either a binary classification (artifact prese…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1291· cite & verify
Slice-Wise Quality Assessment of High b Value Breast DWI via Deep Learning-Based Artifact Detection
TORNAI-SpectrumOfMedicalImaging by Martin Tornai. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers developed convolutional neural networks to automatically flag intensity artifacts on high b-value breast diffusion-weighted MRI. Using 11,806 slices from 156 exams, DenseNet121 distinguished hyper- and hypointense artifacts with AUROCs of 0.91 and 0.95 on binary tasks and weighted AUROCs of 0.84 and 0.90 on multiclass tasks.

Reliable slice-wise quality assessment could help prevent artifact-related misinterpretation as DWI is added to breast MRI protocols to support lesion detection. The work remains retrospective and limited to one dataset, so prospective testing across scanners, protocols, and patient populations is needed to confirm generalizability and clinical impact.

Main points

  1. Retrospective study used 11,806 DWI slices from 156 examinations of female patients between 2022 and mid-2023 at 3T with b values 50, 750, and 1500 s/mm2.
  2. Three CNNs (DenseNet121, ResNet18, SEResNet50) were trained for binary artifact presence; best model DenseNet121 was tested on independent holdout sets of 1836 and 1822 slices.
  3. Multiclass classification yielded weighted AUROCs of 0.84 and 0.90, with Grad-CAM bounding boxes scored 3.52 ± 0.81 and 3.62 ± 0.65 on 5-point Likert-like scale by radiologist.

The gain

A DenseNet121 model trained on 11,806 breast DWI slices can automatically flag hyper- and hypointense artifacts on high b-value images with high discrimination, supporting quality assessment for lesion detection.

The rundown

The team labeled slices for hyper- and hypointense artifacts and split them into training, validation, and holdout sets (8164/1806/1836 and 8164/1820/1822). They compared DenseNet121, ResNet18, and SEResNet50 for binary detection, then extended the best model to multiclass intensity classification.

Evaluation used accuracy, precision, recall, AUROC and AUPRC, plus radiologist review of Grad-CAM heatmap bounding boxes on 150 validation and 200 holdout slices. Results were reported on single-slice high b-value diffusion-weighted images at b = 1500 s/mm2.

Sources

  1. Peer-reviewedJournal of Magnetic Resonance Imaging2026-10-05

The debate