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.
Figures & tables
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 binary classification of hyper- and hypointense artifacts. The best performing model was used for inference on the test set for binary classification and subsequently retrained for multiclass classification of hyper- and hypointense artifacts.
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 hypointense artifacts.
Network
Classification
Learning Rate
Hyperintense Artifact
DenseNet121
Binary
8×10−6
ResNet18
Binary
4×10−6
SEResNet50
Binary
8×10−6
DenseNet121
Multiclass
8×10−6
Hypointense Artifact
Table 1: Summary of experiments conducted, showing the network architectures, configurations, and learning rates used for classifying hyper- and hypointense artifacts.
Model
Accuracy
Precision
Recall
AUROC
AUPRC
Hyperintense Artifacts (n = 1836 slices)
DenseNet121
0.84
0.60
0.82
0.92
0.77
Hypointense Artifacts (n = 1822 slices)
DenseNet121
0.85
0.76
0.91
0.94
0.92
Table 2: Quantitative performance metrics (Accuracy, Precision, Recall, area under receiver operating characteristics curve (AUROC), and area under precision recall curve (AUPRC)) for binary hyper- and hypointense artifact classification using DenseNet121 individually on the holdout test set.
Figure 3: True Positive (TP), False Positive (FP), and False Negative (FN) cases of bounding boxes predicted on masked high b-value (b = 1500 s/mm 2 ) DWI slices for binary hyper- and hypointense artifact classification on the holdout test set using DenseNet121. In FN cases, the red arrow shows the correct region of artifact.
Model
Accuracy
Precision
Recall
AUROC
AUPRC
Hyperintense Artifacts (n = 1836 slices)
DenseNet121
0.59
0.76
0.59
0.85
0.75
Hypointense Artifacts (n = 1822 slices)
DenseNet121
0.61
0.68
0.61
0.88
0.69
Table 3: Weighted quantitative performance metrics (Accuracy, Precision, Recall, area under receiver operating characteristics curve (AUROC), and area under precision recall curve (AUPRC)) when predicted for multiclass hyper- and hypointense artifact classification using DenseNet121 on the respective holdout test sets.
Figure 4: Class-wise confusion matrices and area under receiver operating characteristics curves (AUROCs) illustrating the performance of the DenseNet121 model for multiclass classification of hyper- and hypointense artifacts on their respective holdout test sets. a) Confusion matrix and AUROC for hyperintense multiclass classification. b) Confusion matrix and AUROC for hypointense multiclass classification.
Figure 5: The heatmap shows the agreement between ground truth (GT), Validator 1 (V1), and Validator 2 (V2) and the classes predicted by DenseNet121 on the holdout test set using Cohen’s kappa score for hyper- and hypointense artifact obtained from multiclass classification.
Sequence
TR (ms)
TE (ms)
TI (ms)
Matrix Size
Slice Thickness (mm)
DWI acquired by Skyra-Fit
7030–8360
66
250
256–264 × 160–208 × 37–45
4
DWI acquired by MAGNETOM VIDA
5810–13440
66–80
220–250
256–264 × 160–208 × 37–45
4
Supplemental Table S1: Summary of diffusion weighted imaging (DWI) acquisition parameters including repetition time (TR), echo time (TE), inversion time (TI), matrix size and slice thickness for the dataset used in our study.
Model
Accuracy
Precision
Recall
AUROC
AUPRC
Hyperintense Artifacts (n = 1806 slices)
DenseNet121
0.81
0.58
0.81
0.89
0.73
ResNet18
0.80
0.56
0.79
0.87
0.72
SEResNet50
0.79
0.55
0.74
0.84
0.66
Hypointense Artifacts (n = 1820 slices)
DenseNet121
0.82
0.72
0.88
0.92
0.88
Supplemental Table S2: The table shows the quantitative performance metrics (Accuracy, Precision, Recall, area under receiver operating characteristics curve (AUROC), and area under precision recall curve (AUPRC)) for binary hyper- and hypointense artifact classification using last-epoch checkpoints of DenseNet121, ResNet18, and SEResNet50 on the validation set.
Supplemental Figure S1: True Positive (TP), False Positive (FP), and False Negative (FN) cases of bounding boxes predicted on masked high b-value (b = 1500 s/mm 2 ) DWI slices for binary hyper- and hypointense artifact classification on the validation set using DenseNet121, ResNet18, and SEResNet50. In FN cases, the red arrow shows the correct region of artifact.
Supplemental Figure S2: Bounding box quality scores for scoring the ability to enclose an artifact on a 5-point Likert-like scale for DenseNet121, ResNet18, and SEResNet50 are shown in percentages on n = 200 validation set hyper- and hypointense artifact binary classification cases.
GT
V1
V2
DenseNet121
Hyperintense Artifacts
GT
x
0.34 (95% CI: 0.31, 0.37)
0.18 (95% CI: 0.15, 0.21)
0.33 (95% CI: 0.3, 0.36)
V1
x
x
0.13 (95% CI: 0.11, 0.16)
0.39 (95% CI: 0.36, 0.42)
V2
x
x
x
0.12 (95% CI: 0.09, 0.15)
Hypointense Artifacts
GT
x
0.44 (95% CI: 0.42, 0.47)
0.25 (95% CI: 0.22, 0.28)
0.46 (95% CI: 0.43, 0.48)
Supplemental Table S3: Cohen’s kappa with 95% confidence interval (CI) agreement between ground truth (GT), Validator 1 (V1), Validator 2 (V2), and DenseNet121 predicted classes for hyper- and hypointense artifacts on the holdout test set.
Computer Sciences Department, University of Oviedo, Asturias, Spain · Biomedical Engineering Center (BME), University of Oviedo, Asturias, Spain · Electrical Engineering Department, University of Oviedo, Asturias, Spain +6
Institue for Applied Medical Informatics, University Medical Center Hamburg-Eppendorf, Hamburg, 20246, Germany · Institute of Computational Neuroscience, University Medical Center Hamburg-Eppendorf, 20246, Hamburg, Germany · Center for Biomedical Artificial Intelligence (bAIome), University Medical Center Hamburg-Eppendorf, 20246, Hamburg, Germany