Slice‐Wise Quality Assessment of High b Value Breast DWI via Deep Learning‐Based Artifact Detection
Abstract
Background
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.
Purpose
) using deep learning, employing either a binary classification (artifact presence) or a multiclass classification (artifact intensity) approach on a slice-wise dataset. STUDY TYPE: Retrospective. STUDY POPULATION: 11,806 DWI slices were acquired from 156 examinations performed on female patients between 2022 and mid-2023. For both hyper- and hypointense artifact classification, the slices were split into test, validation, and holdout test set with 8164/1806/1836 and 8164/1820/1822 slices in these sets respectively. FIELD STRENGTH/SEQUENCE: . ASSESSMENT: Three convolutional neural network (CNN) architectures (DenseNet121, ResNet18, and SEResNet50) were trained for binary classification of hyper- and hypointense artifacts on slice-wise labeled data. The best performing model (DenseNet121) was applied to an independent holdout test set and further trained separately for multiclass classification. STATISTICAL TESTS: Performance of the networks was evaluated using accuracy, precision, recall, and areas under the receiver operating characteristic curve (AUROC) and under the precision recall curve (AUPRC). Radiologist evaluated bounding box positions on a 5-point Likert-like scale across 150 slices for validation and 200 slices for the holdout test set, derived from the network's Grad-CAM heatmaps.
Results
DenseNet121 achieved AUROCs of 0.91 and 0.95 for hyper- and hypointense artifact detection, respectively, and weighted AUROCs of 0.84 and 0.90 for multiclass classification on single-slice high b value diffusion-weighted images, achieving mean scores of 3.52 ± 0.81 for hyperintense artifacts and 3.62 ± 0.65 for hypointense artifacts. DATA
Conclusion
) using CNNs, particularly DenseNet121, suggests potential reliability and requires further validation. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 3.