3D Masked Autoencoders with Application to

Anomaly Detection in Non-Contrast Enhanced

Breast MRI

Daniel M. Lang1,2 , Eli Schwartz3,4 , Cosmin I. Bercea1,2 , Raja Giryes4 , and

Julia A. Schnabel1,2,5
arXiv:2303.05861v1 [eess.IV] 10 Mar 2023

[email protected]
1
Helmholtz Munich, Germany
2
Technical University of Munich, Germany
3
IBM Research AI, Israel
4
Tel-Aviv University, Israel
5
King’s College London, United Kingdom

Abstract. Self-supervised models allow (pre-)training on unlabeled data

and therefore have the potential to overcome the need for large anno-
tated cohorts. One leading self-supervised model is the masked autoen-
coder (MAE) which was developed on natural imaging data. The MAE
is masking out a high fraction of visual transformer (ViT) input patches,
to then recover the uncorrupted images as a pretraining task. In this
work, we extend MAE to perform anomaly detection on breast magnetic
resonance imaging (MRI). This new model, coined masked autoencoder
for medical imaging (MAEMI) is trained on two non-contrast enhanced
MRI sequences, aiming at lesion detection without the need for intra-
venous injection of contrast media and temporal image acquisition. Dur-
ing training, only non-cancerous images are presented to the model, with
the purpose of localizing anomalous tumor regions during test time. We
use a public dataset for model development. Performance of the architec-
ture is evaluated in reference to subtraction images created from dynamic
contrast enhanced (DCE)-MRI.

Keywords: Anomaly Detection · Masked Autoencoder · Unsupervised

Learning.

1 Introduction
Annotation of medical data requires expert knowledge or labor-intensive test-
ing methods, leading to high curation costs. Therefore, labeled medical imaging
datasets are typically several orders of magnitude smaller than datasets generally
encountered in computer vision. Deep learning networks require large amounts
of data to be trained, making deployment of models in the medical domain cum-
bersome [4]. Self-supervised learning aims at model development in the absence
of labeled examples and has the power to overcome those limiting factors [23].
A pretraining task is utilized to induce prior knowledge into the model, which
2 Lang et al.

will then be fine-tuned for the respective downstream task of interest. One lead-
ing self-supervised approach is the masked autoencoder (MAE) [13], which was
developed on natural imaging data. MAE is a transformer based autoencoder
(AE) model that randomly removes a high fraction of its input patches, with
the intention to recover the uncorrupted images as a self-supervised task.
In addition to constraints in data acquisition, medical datasets are also of-
ten highly imbalanced, featuring a skewed proportion of healthy and unhealthy
examples. Anomaly detection (AD) models are designed to identify rare, uncom-
mon elements that differ significantly from normal cases. In the medical domain,
such models are employed to distinguish abnormal patterns of unhealthy exam-
ples from normal patterns of healthy cases.
Self-supervised anomaly detection combines both training strategies, aiming
to identify abnormal cases without the requirement for labeled examples. This
can be achieved by reconstruction-based methods [3,10]. Models are trained to
recover their input, while restrictions on the architecture are applied. Such re-
strictions can be imposed by information bottlenecks [5] or the alteration of input
images by application of noise [15,27] or removal of image parts [29,28]. During
training, normal examples are shown to the model. In this way, the model is only
able to reconstruct image parts stemming from the normal distribution reason-
ably well, while abnormal image parts result in higher error rates that can be
utilized to generate anomaly maps during test time. Schwarz et al. [21] modified
MAE to perform anomaly detection on natural imaging data. Most AD models
in the area of medical imaging have been developed on MRI of the brain, e.g.
[1,2,15]. Further areas of application include, e.g. chest X-ray, optical coherence
tomography (OCT) and mammography [25].
We aim at model development on breast MRI, which is the most sensitive
breast cancer imaging method [16], applied for tumor staging but also cancer
screening. DCE-MRI refers to the acquisition of images before, during and after
intravenous injection of contrast media, which improves the signal intensity of
neoangiogenically induced vascular changes that allows for better detection of
lesions [26]. However, long scan times and high costs limit widespread use of
the technique, leading different studies to investigate the ability to abbreviate
contrast enhanced breast MRI protocols [16]. We demonstrate the capability of
self-supervised models for anomaly detection on non-contrast enhanced breast
MRI, which reduces the number of required image sequences dramatically and
therefore results in even faster image acquisition. Moreover, no intravenous in-
jection of contrast media is needed, which is known to be able to cause side
effects [12].

Contribution In this work we remodel MAE and extended and further develop
the approach of Schwarz et al. [21], enabling self-supervised anomaly detection
on 3D multi-spectral medical imaging data. To do so, we advance the definition
of input patches and positional embedding of the ViT architecture and refine
the random masking strategy of He et al. [13]. We then train the model on
non-contrast enhanced breast MRI. During training only healthy, non-cancerous
 MAEMI for Anomaly Detection in Breast MRI 3

breast MRIs are shown to the model, aiming to identify breast lesions as anoma-
lies during test time. To the best of our knowledge, we are the first to make the
following contributions:

– We extend and further refine MAEs to perform anomaly detection on 3D

multi-spectral medical imaging.
– We investigate the capability of self-supervised anomaly detection to identify
pathologies in breast MRI.
– We assess the performance of anomaly detection algorithms in reference to
DCE-MRI subtraction images. Paving the way for a more widespread use of
MRI in breast cancer diagnosis.

2 Related Work

The ability of deep convolutional AEs to perform reconstruction based self-

supervised anomaly detection on imaging data has been investigated by several
studies, see e.g. [1,2,22]. Kascenas et al. [15] trained a denoising AE on brain
MRI, such that unhealthy pathologies were removed during test time. Zavrtanik
et al. [29] developed a convolutional AE, masking part of the input to perform
inpainting on natural imaging and video data.
MAE has been developed on 2D natural images, to be finetuned on a classi-
fication problem. In the context of classification the approach has been modified
in several different ways. Feichtenhofer et al. [9] enhanced the model to do clas-
sification on natural video data.
Prabhakar et al. [18] improved the initial MAE architecture, by incorporation
of a contrastive and a auxiliary loss term, to perform classification on brain
MRI. Our model relies on image reconstruction and computation of a voxel-wise
difference in the downstream task. Therefore, we use unaltered mean squared
error (MSE) as a loss.
Due to its high masking ratio, anomalies are likely to be removed by MAE.
Which led Tian et al. [24] to employ the model on anomaly detection in 2D
colonoscopy and X-ray data. They introduced memory-augmented self-attention
and a multi-level cross-attention operator in the underlying ViT architecture,
to limit dependency on random masking. In contrast, we train our model on
multi-spectral 3D data following the principle strategy of Schwartz et al. [21].
The approach does not rely on any modifications in the ViT architecture, and
maximizes the likelihood for the anomaly to be removed by application of a high
number of random masks. We extend and further develop the model trained
on natural imaging data to be able to handle multi-spectral volumetric medical
imaging data.
MAE based anomaly detection models employ a problem independent pre-
training task, recovering pseudo-normal images from the masked input. Denois-
ing autoencoders (DAEs) used for AD are aiming to achieve the same effect by
pretraining on noise removal. However, selection of noise has to fit the distribu-
tion of possible anomalies for the approach to succeed. Hence, a model trained
4 Lang et al.

on one specific task is very unlikely to succeed reasonably well in another prob-
lem setting. In contrast to that, the ability of MAE based anomaly detection to
succeed in modified settings has been proven by [21], achieving state of the art
(SOTA) performance on few- and zero-shot problems.
Identification of lesions in breast MRI has only been performed by supervised
models so far. Maicas et al. [17] trained a deep Q-network for breast lesion
detection, Ayatollahi et al. modified RetinaNet and Herent et al. [14] utilized
a 2D ResNet50. Notably, all of those approaches were trained on DCE-MRI
data, relying on injection of contrast media. Whereas, we perform self-supervised
anomaly detection on non-contrast enhanced MRI.

3 Dataset

We use the public Duke-Breast-Cancer-MRI cohort [19,20] from The Cancer

Imaging Archive [7]. The set includes axial breast MRI data of 922 patients com-
prising a non-fat saturated pre-contrast T1-weighted sequence, a fat-saturated
T1-weighted pre-contrast sequence and several post-contrast fat-saturated T1-
weighted sequences. Tumor lesion annotations were given in the form of bound-
ing boxes. Furthermore, a U-Net model for generation of breast tissue masks
was provided [6]. All images were standardized to the same voxel spacing of
0.75mm × 0.75mm × 1.0mm, cropped to involve the chest area only, and nor-
malized to a mean and standard deviation value of 0.5 and 0.25 per image. Cases
involving bilateral breast cancer were removed from the cohort. Thus, each of the
remaining patients exhibited one breast that contained a tumor lesion, treated
as abnormal/unhealthy, and another breast not affected by cancer, treated as
normal/healthy. The dataset was split into a training set of 745 patients, a val-
idation set of 50 patients and a test set of 100 patients. Notably, no normal
- abnormal pairs of the same patient were involved in different datasets. Due
to the high memory requirements of transformer models, MRI-patches of size
240 × 168 × 8 voxels in lateral-posterior-superior (LPS) directions were cropped
from both MRI sequences to be then divided into ViT-patches and processed by
the encoder.
Following clinical routine, subtraction images S between the image acquired
before Ipre and all of the m images acquired after injection of contrast media
k
Ipost were computed:
m
1 X k

2
S= Ipre − Ipost ∗ min3×3×2 , (1)
m
k=0

with a minimum filter of size 3 × 3 × 2 applied for noise removal.

4 Method

A scheme of our model can be seen in Figure 1. We modified the ViT architecture
 MAEMI for Anomaly Detection in Breast MRI 5

Fig. 1. DCE-MRI imaging vs. MAEMI. For DCE-MRI, several MRIs before, dur-
ing and after injection of contrast media are acquired. MAEMI uses different random
masks for generation of pseudo-healthy recovered images. Both methods construct er-
ror maps by calculation of the mean squared error difference between each of the
post-contrast/reconstructed images with the pre-contrast/ uncorrupted image.

of [8] to 3D multispectral MRI, i.e. the positional embedding and definition of

ViT-patches was redefined and enhanced to incorporate a third dimension. This
3D-ViT was then embedded in the MAE approach of [13]. To do so, generation
of mask tokens and random masking of ViT-patches had to be remodeled. The
approach coined masked autoencoder for medical imaging (MAEMI) uses a 3D-
ViT with 12 transformer blocks and a embedding dimension of 768 as encoder
while for the decoder a embedding dimension of 384 and a depth of 4 has been
chosen.
In addition, we further improved the anomaly detection model of [21]. MRIs
were processed by patches, such that memory requirements of the transformer
based architecture could be reduced, enabling exploitation of whole MRI vol-
umes. A overlapping patch scheme has been chosen, in order to reduce artifacts
on patch borders. Anomaly maps were generated from two input sequences, i.e.
non-fat saturated (NFS) and fat saturated (FS).
Error maps per input sequence Eiseq were constructed by the MSE between
the reconstructed patches Riseq and the unmasked MRI-patch I seq :

Eiseq = (I seq − Riseq )2 ∗ min3×3×2 , (2)

with a minimum filter of size 3 × 3 × 2 applied for further reduction of artifacts.

On a voxel level, final error scores were computed by the mean value of all patch
predictions. Error maps of both multi-spectral input sequences, NFS and FS,
6 Lang et al.

were then summed up and convolved with the same minimum filter as before:
1
E NFS + E FS ∗ min3×3×2 ,


E= (3)
2
for generation of a final MR image level anomaly map.

Training Specifics During training, only MRI-patches of healthy breasts, con-

taining no tumor lesions, were shown to the model with patches being cropped
randomly. In addition to random cropping, random flipping on the coronal and
sagittal plane was applied as a augmentation technique during training. A batch
size of 6 and a learning rate of 10−3 were applied. Each model was trained for
1000 epochs, with the number of warm up epochs [11] set to 7. Weights of the
trained model of [13], developed on ImageNet, were used to initialize the trans-
former layers in the encoder, while weights of the encoding layer and the decoder
were randomly initialized. During test time a stride of size 64×42×2 performing
6 repetitions was used to process whole MRI volumes.

Metrics We used voxel wise area under the receiver operating characteristics
curve (AUROC) and average precision (AP) as performance measures. Only
voxels lying inside the breast tissue segmentation mask were taken into account
for computation, as injection of contrast media leads also to an uptake in tissue
lying outside the breast area, ref. Figure 5 the Supplemental Materials. However,
for AP one has to consider the large imbalance between normal and abnormal tis-
sue labels, leading to an expected small baseline performance. Moreover, ground
truth annotations were only given in the form of bounding boxes, depicting only
a rough delineation of tumor tissue with several ground truth true positive (TP)
scores involved that should in fact be true negative (TN). This has an higher
impact on AP than on AUROC, as TP scores are involved in precision and recall
but not in the false positive rate (FPR) of the ROC, which also takes TN labels
into account.

5 Results
ViT-patch size and masking ratio have been varied for hyperparameter tuning.
The best performing model featured a masking ratio of 90% and a ViT-patch
size of 8 × 8 × 2, AUROC and AP results are shown in Table 1. Example results
are shown in Figure 2. Mean baseline performance of the AP measure, given by
the number of voxels inside the bounding box divided by the number of voxels
lying inside the breast tissue segmentation mask, was given by 0.046.

Ablation Studies We studied the influence of the patch size and masking ratio
on model performance. Figure 3 presents the dependency of AUROC and AP
on the masking ratio for a fixed ViT-patch size of 8 × 8 × 2. Dependency on
ViT-patch size for a fixed masking ratio of 90% is given in Table 2. The Nvidia
RTX A6000 used for training, featuring 48 GB of memory, only allowed for a
smallest size of 8 × 8 × 2. Therefore, the slice dimension of ViT-patches was fixed
at a value of 4 pixels, probing only different axial sizes.
 MAEMI for Anomaly Detection in Breast MRI 7

Table 2. Ablation study on different (ax-

Table 1. MAEMI achieves a higher ial) ViT-patch sizes, for a fixed masking
AUROC, while DCE-MRI features a ratio of 90%. Smaller patch sizes lead to
higher AP. However, significance of AP higher performance.
results are limited (see Section 4).
ViT-patch size AUROC AP
AUROC AP 6×6×4 0.724 0.0784
MAEMI 0.732 0.081 8×8×4 0.712 0.0777
DCE-MRI 0.705 0.127 12 × 12 × 4 0.660 0.0750
24 × 24 × 4 0.546 0.0560

Fig. 2. Example results. The first two columns show the non-contrast enhanced images
used as an input to the anomaly detection model, and the last two columns present
subtraction images generated by DCE-MRI and anomaly detection maps generated by
MAEMI, respectively. For patients in rows A, B and C, anomaly maps show superior
performance over subtraction images. For patient D, both methods are able to identify
the pathology. For patient E, our model only detects the borders of the pathology,
while the subtraction image identifies the lesion.
 MAEMI for Anomaly Detection in Breast MRI

Fig. 3. Ablation study on the masking ratio for a fixed ViT-patch size of 8 × 8 × 2.
High masking ratios lead to better performance, with an optimum reached at 90%.
Afterwards, performance suffers from a steep decline.

6 Discussion

We developed a new transformer-based AE model that can be trained on multi-

spectral volumetric medical imaging data. We have applied our model to anomaly
detection on non-contrast enhanced breast MRI. Model performance was at the
same level as for DCE-MRI generated subtraction images. Thus, we were able
to demonstrate the general ability for automated identification of suspicious
pathologies on non-contrast enhanced breast MRI.
MAEMI achieved a higher AUROC while DCE-MRI generated subtraction
images resulting in better AP performance. Limitations of evaluation metrics
due to ground truth labels given in the form of bounding boxes were stated in
Section 4. Determination of an optimal performance measure for ground truth
bounding boxes in the case of unsupervised anomaly detection remains an active
area of research.
We found an optimal masking ratio of 90% for our model. Feichtenhofer et
al. [9] identified the same ratio to work best for video classification. However,
masking plays a significantly different role for MAE architectures utilized for
anomaly detection, with ViT-patch removal not only applied during training but
also during test time. Tian et al. [24] employed the standard masking ratio of
75% for their model, trained to identify pathologies on 2D X-ray and colonoscopy
data and did not report any ablation studies.
Automated generation of anomaly maps from non-contrast enhanced imag-
ing allows for identification of suspicious lesions without the need for temporal
imaging, which results in a drastic reduction of costs and acquisition time, two
major factors limiting widespread application of MRI in breast cancer screening
[16]. Therefore, we paved the way for a more widespread application of MRI in
breast cancer diagnosis. Furthermore, patients can potentially be spared from
intravenous injection of contrast media, which is known to be able to cause side
effects [12]. Clinical differences and benefits between MAEMI and DCE-MRI
will still need to be investigated in a larger clinical study.
 Lang et al.

Data Use Declaration All data used for this study is publicly available from
The Cancer Imaging Archive [19,7] under the CC BY-NC 4.0 license.

Supplementary Material

Fig. 4. Reconstruction examples. The left block shows axial slices of T1 non-fat satu-
rated MRI-patches and the right block T1 fat saturated slices. The first column shows
unaltered MRI-patches, the second column the masked model input and the third col-
umn the MRI-patches recovered by MAEMI. Examples represent a masking ratio of
90% (for the whole 3D patch) and a ViT-patch size of 8 × 8 × 2.
 MAEMI for Anomaly Detection in Breast MRI

Fig. 5. Subtraction images and anomaly maps were multiplied with segmentation
masks to remove anomalies lying obviously outside of the breast tissue. This is mainly
needed as contrast agent is also taken up in organs outside the breast. The left column
shows the raw subtraction/anomaly map, and the right column the raw maps multi-
plied with the segmentation mask of the image in the upper left corner. Performance
metrics were only calculated for voxels lying inside the segmentation mask, limiting
the influence of trivial predictions, i.e. voxels that represent air do not containing any
anomalies.
 Lang et al.

References
1. Baur, C., Denner, S., Wiestler, B., Navab, N., Albarqouni, S.: Autoencoders for
unsupervised anomaly segmentation in brain MR images: a comparative study.
Medical Image Analysis 69, 101952 (2021)
2. Bercea, C.I., Wiestler, B., Rueckert, D., Albarqouni, S.: Federated disentangled
representation learning for unsupervised brain anomaly detection. Nature Machine
Intelligence 4(8), 685–695 (2022)
3. Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsuper-
vised defect segmentation by applying structural similarity to autoencoders. arXiv
preprint arXiv:1807.02011 (2018)
4. Ching, T., Himmelstein, D.S., Beaulieu-Jones, B.K., Kalinin, A.A., Do, B.T., Way,
G.P., Ferrero, E., Agapow, P.M., Zietz, M., Hoffman, M.M., et al.: Opportunities
and obstacles for deep learning in biology and medicine. Journal of The Royal
Society Interface 15(141), 20170387 (2018)
5. Chow, J.K., Su, Z., Wu, J., Tan, P.S., Mao, X., Wang, Y.H.: Anomaly detection
of defects on concrete structures with the convolutional autoencoder. Advanced
Engineering Informatics 45, 101105 (2020)
6. Chris, L.: 3D-Breast-FGT-and-Blood-Vessel-Segmentation. https://github.com/
mazurowski-lab/3D-Breast-FGT-and-Blood-Vessel-Segmentation (2022)
7. Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J., Koppel, P., Moore, S.,
Phillips, S., Maffitt, D., Pringle, M., et al.: The Cancer Imaging Archive (TCIA):
maintaining and operating a public information repository. Journal of digital imag-
ing 26, 1045–1057 (2013)
8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner,
T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is
worth 16x16 words: Transformers for image recognition at scale. arXiv preprint
arXiv:2010.11929 (2020)
9. Feichtenhofer, C., Fan, H., Li, Y., He, K.: Masked autoencoders as spatiotemporal
learners. arXiv preprint arXiv:2205.09113 (2022)
10. Gong, D., Liu, L., Le, V., Saha, B., Mansour, M.R., Venkatesh, S., Hengel, A.v.d.:
Memorizing normality to detect anomaly: Memory-augmented deep autoencoder
for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF Interna-
tional Conference on Computer Vision. pp. 1705–1714 (2019)
11. Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tul-
loch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in 1
hour. arXiv preprint arXiv:1706.02677 (2017)
12. Hasebroock, K.M., Serkova, N.J.: Toxicity of MRI and CT contrast agents. Expert
opinion on drug metabolism & toxicology 5(4), 403–416 (2009)
13. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are
scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition. pp. 16000–16009 (2022)
14. Herent, P., Schmauch, B., Jehanno, P., Dehaene, O., Saillard, C., Balleyguier, C.,
Arfi-Rouche, J., Jégou, S.: Detection and characterization of MRI breast lesions
using deep learning. Diagnostic and interventional imaging 100(4), 219–225 (2019)
15. Kascenas, A., Pugeault, N., O’Neil, A.Q.: Denoising autoencoders for unsupervised
anomaly detection in brain MRI. In: International Conference on Medical Imaging
with Deep Learning. pp. 653–664. PMLR (2022)
16. Leithner, D., Moy, L., Morris, E.A., Marino, M.A., Helbich, T.H., Pinker, K.: Ab-
breviated MRI of the breast: does it provide value? Journal of Magnetic Resonance
Imaging 49(7), e85–e100 (2019)
 MAEMI for Anomaly Detection in Breast MRI

17. Maicas, G., Carneiro, G., Bradley, A.P., Nascimento, J.C., Reid, I.: Deep reinforce-
ment learning for active breast lesion detection from DCE-MRI. In: Medical Image
Computing and Computer Assisted Intervention- MICCAI 2017: 20th International
Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part
III. pp. 665–673. Springer (2017)
18. Prabhakar, C., Li, H.B., Yang, J., Shit, S., Wiestler, B., Menze, B.: ViT-AE++: Im-
proving Vision Transformer Autoencoder for Self-supervised Medical Image Rep-
resentations. arXiv preprint arXiv:2301.07382 (2023)
19. Saha, A., Harowicz, M., Grimm, L., Weng, J., Cain, E., Kim, C., Ghate, S., Walsh,
R., Mazurowski, M.: Dynamic contrast-enhanced magnetic resonance images of
breast cancer patients with tumor locations. The Cancer Imaging Archive (2021)
20. Saha, A., Harowicz, M.R., Grimm, L.J., Kim, C.E., Ghate, S.V., Walsh, R.,
Mazurowski, M.A.: A machine learning approach to radiogenomics of breast can-
cer: a study of 922 subjects and 529 DCE-MRI features. British journal of cancer
119(4), 508–516 (2018)
21. Schwartz, E., Arbelle, A., Karlinsky, L., Harary, S., Scheidegger, F., Doveh, S.,
Giryes, R.: MAEDAY: MAE for few and zero shot AnomalY-Detection. arXiv
preprint arXiv:2211.14307 (2022)
22. Somepalli, G., Wu, Y., Balaji, Y., Vinzamuri, B., Feizi, S.: Unsupervised anomaly
detection with adversarial mirrored autoencoders. In: Uncertainty in Artificial In-
telligence. pp. 365–375. PMLR (2021)
23. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness
of data in deep learning era. In: Proceedings of the IEEE international conference
on computer vision. pp. 843–852 (2017)
24. Tian, Y., Pang, G., Liu, Y., Wang, C., Chen, Y., Liu, F., Singh, R., Verjans,
J.W., Carneiro, G.: Unsupervised anomaly detection in medical images with
a memory-augmented multi-level cross-attentional masked autoencoder. arXiv
preprint arXiv:2203.11725 (2022)
25. Tschuchnig, M.E., Gadermayr, M.: Anomaly detection in medical imaging-a mini
review. In: Data Science–Analytics and Applications: Proceedings of the 4th In-
ternational Data Science Conference–iDSC2021. pp. 33–38. Springer (2022)
26. Turnbull, L.W.: Dynamic contrast-enhanced MRI in the diagnosis and management
of breast cancer. NMR in Biomedicine: An International Journal Devoted to the
Development and Application of Magnetic Resonance In Vivo 22(1), 28–39 (2009)
27. Wyatt, J., Leach, A., Schmon, S.M., Willcocks, C.G.: AnoDDPM: Anomaly De-
tection With Denoising Diffusion Probabilistic Models Using Simplex Noise. In:
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog-
nition (CVPR) Workshops. pp. 650–656 (June 2022)
28. Yan, X., Zhang, H., Xu, X., Hu, X., Heng, P.A.: Learning semantic context from
normal samples for unsupervised anomaly detection. In: Proceedings of the AAAI
Conference on Artificial Intelligence. vol. 35, pp. 3110–3118 (2021)
29. Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual
anomaly detection. Pattern Recognition 112, 107706 (2021)