[論文レビュー] Slice-wise quality assessment of high b-value breast DWI via deep learning-based artifact detection

Ameya Markale, Luise Brock+11 more|arXiv (Cornell University)|Mar 4, 2026

MRI in cancer diagnosis被引用数 0

ひとこと要約

The study uses CNNs to detect and classify hyper- and hypointense artifacts on high b-value breast DWI (b=1500 s/mm2) on a slice-wise basis, with DenseNet121 achieving strong AUROC/AUPRC on holdout data and bounding-box visualization via Grad-CAM.

ABSTRACT

Diffusion-weighted imaging (DWI) can support lesion detection and characterization in breast magnetic resonance imaging (MRI), however especially high b-value diffusion-weighted acquisitions can be prone to intensity artifacts that can affect diagnostic image assessment. This study aims to detect both hyper- and hypointense artifacts on high b-value diffusion-weighted images (b=1500 s/mm2) using deep learning, employing either a binary classification (artifact presence) or a multiclass classification (artifact intensity) approach on a slice-wise dataset.This IRB-approved retrospective study used the single-center dataset comprising n=11806 slices from routine 3T breast MRI examinations performed between 2022 and mid-2023. Three convolutional neural network (CNN) architectures (DenseNet121, ResNet18, and SEResNet50) were trained for binary classification of hyper- and hypointense artifacts. The best performing model (DenseNet121) was applied to an independent holdout test set and was further trained separately for multiclass classification. Evaluation included area under receiver operating characteristic curve (AUROC), area under precision recall curve (AUPRC), precision, and recall, as well as analysis of predicted bounding box positions, derived from the network Grad-CAM heatmaps. DenseNet121 achieved AUROCs of 0.92 and 0.94 for hyper- and hypointense artifact detection, respectively, and weighted AUROCs of 0.85 and 0.88 for multiclass classification on single-slice high b-value diffusion-weighted images. A radiologist evaluated bounding box precision on a 1-5 Likert-like scale across 200 slices, achieving mean scores of 3.33+-1.04 for hyperintense artifacts and 2.62+-0.81 for hypointense artifacts. Hyper- and hypointense artifact detection in slice-wise breast DWI MRI dataset (b=1500 s/mm2) using CNNs particularly DenseNet121, seems promising and requires further validation.

研究の動機と目的

• Address the challenge of intensity artifacts in high b-value breast DWI (b = 1500 s/mm2).
• Develop slice-wise CNN-based methods to detect and grade hyper- and hypointense artifacts.
• Evaluate binary artifact presence and multiclass artifact intensity classification.
• Provide localization of artifacts using Grad-CAM derived bounding boxes for explainability.

提案手法

• Construct slice-wise datasets from a single-center, IRB-approved cohort of 11,806 slices from 3T breast MRI with high b-value DWI.
• Compare DenseNet121, ResNet18, and SEResNet50 architectures for binary classification of hyper- and hypointense artifacts; select best model for holdout testing.
• Train the best model for multiclass artifact classification (5-point Likert-like scale retained as classes).
• Apply breast masks to focus training on breast tissue; use data augmentation (rotations, flips) and Adam optimization with class-balancing.
• Use Grad-CAM to generate heatmaps, threshold to 0.2, and derive bounding boxes to localize artifacts.

Figure 1: Schematic overview of the study workflow, including maximum intensity projection (MIP)-based case pre-selection, slice-wise data generation to form ground truth (GT), data splitting, and model selection using DenseNet121, ResNet18, and SEResNet50 architectures on the validation set for bin

実験結果

リサーチクエスチョン

• RQ1 Can CNNs accurately detect hyper- and hypointense artifacts on single-slice high-b-value breast DWI?
• RQ2 Which CNN architecture best detects these artifacts, and does it generalize to an independent holdout set?
• RQ3 How well can the model perform multiclass artifact severity classification?
• RQ4 Can Grad-CAM-based localization provide reliable bounding boxes for artifact regions?

主な発見

• DenseNet121 achieved AUROC 0.92 and AUPRC 0.77 for hyperintense artifact detection on the holdout set.
• DenseNet121 achieved AUROC 0.94 and AUPRC 0.92 for hypointense artifact detection on the holdout set.
• For multiclass hyperintense artifacts, DenseNet121 achieved AUROC 0.85 and AUPRC 0.75 (holdout).
• For multiclass hypointense artifacts, DenseNet121 achieved AUROC 0.88 and AUPRC 0.69 (holdout).
• Bounding boxes based on Grad-CAM had mean localization quality 3.33±1.04 (hyperintense) and 2.62±0.81 (hypointense) on a 1–5 scale.
• Reader–model agreement showed varying Cohen’s kappa across hyper- and hypointense artifact evaluations, indicating some label variability with multiclass predictions.

Figure 2: Slices affected by hyper- and hypointense artifacts on DWI (b = 1500 s/mm 2 ): a) the enclosed region shows hyperintense artifact caused by surface coil flare; b) the enclosed region shows hyperintense artifact that is likely caused by skin folding; (c-d) enclosed regions depict hypointens