Semi-Supervised Anomaly Detection in Skin Lesion Images

Alina Burgert

, Babette Dellen

, Uwe Jaekel

1 a

and Dietrich Paulus

2 b

Faculty of Mathematics, Informatics, and Technology, University of Applied Sciences Koblenz, Joseph-Rovan-Allee 2,

53424 Remagen, Germany

Institute for Computational Visualistics, University Koblenz, Universit

atsstraße 1, 56070 Koblenz, Germany

{burgert, dellen, jaekel}@hs-koblenz.de, paulus@uni-koblenz.de

Keywords:

Anomaly Detection, Semi-Supervised Learning, Dermatology.

Abstract:

Semi-supervised anomaly detection is the task of learning the pattern of normal samples and identifying devi-

ations from this pattern as anomalies. This approach is especially helpful in the medical domain, since healthy

samples are usually easy to collect and time-intensive annotation of training data is not necessary. In derma-

tology the utilization of this approach is not fully explored yet, since most work is limited to cancer detection,

with the normal samples being nevi. This study, instead, investigates the use of semi-supervised anomaly

detection methods for skin disease detection and localization. Due to the absence of a benchmark dataset a

custom dataset was created. Based on this dataset two different models, SimpleNet and an autoencoder, were

trained on healthy skin images only. Our experiment shows that both models are able to distinguish between

normal and abnormal samples of the test dataset, with SimpleNet achieving an AUROC score of 97 % and

the autoencoder a score of 93 %, demonstrating the potential of anomaly detection for dermatological appli-

cations. A visual analysis of corresponding anomaly maps revealed that both models have their own strengths

and weaknesses when localizing the abnormal regions.

1 INTRODUCTION

Over the last decade, deep learning methods have rev-

olutionized diagnostic capabilities in various medi-

cal domains, including dermatology. According to

Chan et al. (2020), the most common machine learn-

ing paradigm used in dermatology is supervised learn-

ing. However, supervised learning requires large an-

notated training datasets. The annotation procedure is

time-intensive and requires the expertise of medical

professionals and can introduce human bias.

An alternative approach, which addresses some of

these drawbacks, is semi-supervised anomaly detec-

tion, as deﬁned by Chandola et al. (2009). Trans-

ferred to the medical domain, the basic idea is learn-

ing the appearance of a healthy state in order to be

able to identify pathologic cases as deviations from

this state. During training, only images showing the

healthy state are required. This is particularly useful

in cases where no or little pathological data is avail-

able and unknown pathologies also need to be recog-

nized (e. g. in the case of rare diseases). In contrast

to supervised learning, no ground truth is required for

https://orcid.org/0000-0002-4275-1430

https://orcid.org/0000-0002-2967-5277

training, which saves valuable time of medical ex-

perts. A limitation of the approach is that it does not

yield a speciﬁc diagnosis.

In medical imaging, anomaly detection is applied

predominantly to brain MRIs (Tschuchnig and Gader-

mayr, 2022). Only a few works on semi-supervised

anomaly detection in dermatology exist. Most of

them aim to detect skin cancer by learning what

normal pigmented skin lesions (nevi) look like (Lu

and Xu, 2018; Zhang et al., 2022; Grignafﬁni et al.,

2023; Cai et al., 2024). In contrast, other studies

learn the appearance of healthy skin without any le-

sions. However, these studies are limited to the detec-

tion of pigmented skin lesions in dermoscopic images

(Shen et al., 2020) and the detection of hand ekzema

(Gonzalez-Jimenez et al., 2023).

The aim of this work is to explore the approach of

semi-supervised anomaly detection for general skin

lesion detection and localization. Due to a lack of

benchmark datasets, we create our own skin anomaly

detection dataset. Based on this dataset, we com-

pare two anomaly detection methods, SimpleNet (Liu

et al., 2023) and a convolutional autoencoder, which

is often used as a baseline method in anomaly detec-

tion.

Burgert, A., Dellen, B., Jaekel, U. and Paulus, D.

Semi-Supervised Anomaly Detection in Skin Lesion Images.

DOI: 10.5220/0013305400003912

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages

535-541

ISBN: 978-989-758-728-3; ISSN: 2184-4321

535

2 STATE OF THE ART

According to Cai et al. (2024) anomaly detection

methods can be categorized into methods based on

reconstruction, self-supervised learning, and feature

reference. Reconstruction-based methods rely on

generative models, e.g., autoencoder, variational au-

toencoder, generative adversarial networks or diffu-

sion models, that are trained to reconstruct healthy

images. When reconstructing abnormal images, it is

assumed that a comparatively large reconstruction er-

ror occurs, which can be interpreted as an anomaly

score. In self-supervised learning, models are trained

on pretext tasks with generated pseudo labels. The

basic idea is that knowledge which is obtained in the

pretext task can be transferred to the anomaly detec-

tion task. Feature-reference-based methods are based

on the disparity between current and reference fea-

tures. For example, a pretrained network can be uti-

lized to extract and save features of normal images in

a memory bank for reference. During inference, fea-

tures of interest are compared to reference features in

order to detect anomalies.

Anomaly detection has also been applied in der-

matology. Lu and Xu (2018), Zhang et al. (2022),

Grignafﬁni et al. (2023) and Cai et al. (2024) utilize

anomaly detection for melanoma detection in dermo-

scopic images, with the normal condition being de-

ﬁned as nevi. For example in Lu and Xu (2018), a

VAE is trained on images of nevi from the ISIC 2018

dataset. Skin diseases such as melanoma or actinic

keratosis are recognized as an anomaly with an AU-

ROC of 0.779 using a reconstruction-based approach.

In contrast to the studies above, the following

works try to detect skin lesions by learning the healthy

appearance of skin without any lesions. Shen et al.

(2020) propose a new method called adGAN for

anomaly detection. In contrast to existing GAN-based

methods, adGAN does not rely on a reconstruction er-

ror for anomaly detection. Instead, the authors follow

a discriminative approach, where fake images gen-

erated from a GAN are used as an abnormal class

and a discriminator model is trained to discriminate

between the normal and the generated abnormal im-

ages. The proposed model is tested on three datasets

including ISIC 2016 to evaluate the performance of

the model in skin lesion detection, where it achieves

an AUC value of 0.98. In Gonzalez-Jimenez et al.

(2023), a score-based diffusion model is used to de-

tect and localize hand eczema. For this purpose,

the diffusion model is trained with images of healthy

hands. The log-likelihood gradient map, which is

analysed at the beginning of the diffusion process,

is used to detect anomalies. At inference time, it is

Table 1: Number of normal and anomalous images by

source dataset used in this study.

Source # Normal # Abnormal

ISIC Archive 160 11

SD-198 11 158

ArsenicSkinImagesBD 175 -

Google Image Search - 21

All 346 190

therefore not necessary to run through the entire time-

consuming and computationally expensive diffusion

process. A test on a private dataset from a university

hospital demonstrates that hand eczema is recognized

with an AUROC of 0.912.

3 METHODS

3.1 Dataset

Since no publicly available dermatological dataset

was suitable for our anomaly detection study, we cre-

ated a custom dataset. This process involved collect-

ing two types of image classes: normal images show-

ing healthy skin to allow the model to learn the ap-

pearance of healthy skin, and abnormal images show-

ing different types of skin pathologies or irregulari-

ties to evaluate the model’s ability to detect anoma-

lies. For a skin image to be classiﬁed as healthy,

it had to show no lesions, erythema or other visi-

ble pathological signs. An exception was made for

pigmented skin lesions, since the study’s focus is

not on skin cancer detection. Pathological images

were selected to represent a variety of anomaly types

including for example erythema, psoriasis, eczema,

hematoma, scars and imprints of clothing. Only stan-

dard clinical photographs were included, while mi-

croscopic and dermoscopic images were excluded to

maintain consistency, because it could be more chal-

lenging for the model to learn generalized patterns

and appearances of skin across different zoom lev-

els. Images of certain body regions, such as hands,

feet, face, and head, were excluded because of their

unique anatomical features and variability in appear-

ance. To further simplify the task and avoid misclas-

siﬁcation of background pixels as anomalies, images

were cropped to exclude non-skin areas. Included im-

ages vary across different factors such as lightning,

skin tone, age, presence of skin folds and body hair

etc. The ﬁnal dataset was created by collecting im-

ages from the following publicly available sources:

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

536

ISIC Archive

: The ISIC Archive, hosted by the In-

ternational Skin Imaging Collaboration (ISIC), con-

tains a large publicly available collection of skin im-

ages. The majority are dermoscopic images of pig-

mented skin lesions, which are not suitable for our

dataset. Instead we ﬁltered the archive images for to-

tal body photographs (TBPs), which yielded 36 im-

ages, showing the posterior torso. Based on these im-

ages, we extracted multiple smaller images of compa-

rable sizes. This resulted in 160 normal images and 11

abnormal images showing scars or imprints of cloth-

ing.

ArsenicSkinImagesBD

: The ArsenicSkinIm-

agesBD dataset (Emu et al., 2024) contains 741

images of 37 arsenic-affected and 741 images of 76

non-arsenic-affected individuals from Bangladesh,

captured by smartphone cameras. Of the 741 non-

affected images, 175 were used as normal images.

The remaining images were excluded due to different

reasons (e.g. duplicates, showing hands / ﬁngers or

potential skin conditions).

SD-198

: SD-198 (Sun et al., 2016) is a benchmark

dataset for clinical skin diseases containing 6,584 im-

ages from 198 classes. We selected 158 images from

the classes acne vulgaris, allergic contact dermatitis,

eczema, erythema annulare centrifugum, erythema

multiforme, factitial dermatitis, guttate psoriaris, pso-

riaris, tinea corporis and used them as abnormal im-

ages. In addition 11 healthy skin patches were ex-

tracted and added to the normal image dataset.

Google Image Search: Another 21 images contain-

ing erythema or hematoma were collected using a

Google Image Search and added to the abnormal

dataset.

Table 1 shows the number of normal and abnormal

images by source dataset. In total 346 normal and 190

anomalous images were collected. The dataset was

splitted into three datasets for training, validation and

evaluation. Models were trained on 250 normal im-

ages. A validation set of 62 images (20 normal and

42 abnormal) was utilized to optimize hyperparame-

ters and to save the best model for evaluation. The

test set for ﬁnal evaluation contains 224 images (76

normal and 148 abnormal).

For the autoencoder, all images were resized to

128 × 128 and pixel values were scaled into a range

of [0, 1]. For SimpleNet, all images were resized to

224 × 224 and pixel values were ﬁrst scaled into a

range of [0, 1] and then normalized according to the

mean and standard deviation of ImageNet as in Liu

https://www.isic-archive.com/

https://data.mendeley.com/datasets/x4hgnjj5gv/2

https://huggingface.co/datasets/

resyhgerwshshgdfghsdfgh/SD-198

et al. (2023). No data augmentation was applied.

3.2 Model Architectures, Training and

Evaluation

In the following we describe the two anomaly detec-

tion models, the training procedure and the evaluation

metrics used in this study.

SimpleNet: SimpleNet was introduced by Liu et al.

(2023) for the task of detecting and localizing anoma-

lies in industrial images. The authors argue that ex-

isting approaches (e. g. reconstruction- and feature-

based) have some drawbacks and therefore proposed

SimpleNet which combines several approaches and

comes with further improvements. SimpleNet con-

sists of four components. The ﬁrst component is the

feature extractor, a pretrained neural network used for

extracting local image features. Since pretrained net-

works are usually trained on natural images such as

ImageNet and not on industrial or medical images, a

simple neural network called feature adaptor is uti-

lized to map the extracted features into the target do-

main. The third component is an anomalous feature

generator which artiﬁcially generates anomalous fea-

tures by adding random gaussian noise to normal fea-

tures. Last, a simple discriminator network is trained

to discriminate the normal and the artiﬁcially gener-

ated anomalous features. In contrast to Shen et al.

(2020) the discrimination is performed on individ-

ual local feature vectors, not on whole images. Sim-

pleNet with all its components can be trained in an

end-to-end fashion. During inference the generation

of anomalous features is omitted. Local features are

extracted and adapted from the input image and then

mapped to an anomaly score by the discriminator net-

work. Arranging all local anomaly scores in a 2D-

grid yields an anomaly map, highlighting anomalous

areas in the input image. Based on the anomaly map

an image level anomaly score can be computed. In

the original publication of SimpleNet the maximum

anomaly score is used.

For our experiment we used the same hyperparam-

eter conﬁguration as in Liu et al. (2023). We trained

SimpleNet for 160 epochs with a batchsize of 8 and

saved the best model based on validation anomaly de-

tection performance.

Autoencoder: As a baseline model, we implemented

a convolutional autoencoder (AE), consisting of an

encoder and a symmetrical decoder. The encoder

compresses an input image x ∈ R

H×W ×C

into a latent

feature vector z ∈ R

. Based on this feature vector, the

decoder reconstructs the original image. The encoder

consists of four convolutional layers, each downsam-

pling the image resolution to

. The ﬁrst con-

Semi-Supervised Anomaly Detection in Skin Lesion Images

537

Figure 1: Visualization of the compared methods on randomly drawn abnormal test images. The ﬁgure shows the input

abnormal images (1st row), the reconstructed images and corresponding anomaly maps generated by the autoencoder (2nd

and 3rd row) and the anomaly maps generated by SimpleNet (4th row). Warmer colors of the anomaly maps correspond to

higher pixel-level anomaly scores. Color values of anomaly maps generated by one speciﬁc model are directly comparable,

since they follow the same color scale. For visualization purposes, anomaly score outliers were cut off.

Table 2: Performance of autoencoder with different hyper-

parameter conﬁgurations. The best conﬁguration is high-

lighted in bold.

d AUROC % AUPRC %

16 16 90.6 95.4

16 32 90.1 95.0

16 64 90.2 95.6

32 16 90.1 95.7

32 32 90.0 95.5

32 64 91.3 95.7

volutional layer has a width of C

channels. Each

subsequent layer increases the width by a factor of 2.

Convolutional layers are followed by a ReLU activa-

tion. The output of the last convolutional layer is ﬂat-

tened and processed by a fully-connected layer, which

returns the feature vector of length d. A sigmoid func-

tion is used as a last activation in the decoder to ensure

that the output remains in the range of [0, 1]. Basic

width C

and latent dimension d were conﬁgured in

a hyperparameter optimization step by choosing the

model with the best anomaly detection performance

on the validation set (see table 2). As a reconstruction

loss, we used MSE. The model was optimized with

ADAM conﬁgured with an initial learning rate of 1e-

3, a weight decay of 1e-5 and trained for 200 epochs

with a batchsize of 8. The best model (measured in

terms of anomaly detection performance on the vali-

dation set) was updated every epoch and saved after

training.

Evaluation Metrics: Both models yield a real-

Figure 2: Receiver Operating Characteristic Curve, Preci-

sion Recall Curve and corresponding AUC values of au-

toencoder (AE) and SimpleNet (SN).

valued output which can be interpreted as an

anomaly score. Based on this score, we generated

Receiver-Operating-Characteristic- and Precision-

Recall-Curves (ROC and PRC) and calculated the

area under both curves (AUC) to evaluate the capa-

bility of the models to differentiate between normal

and abnormal images. An advantage of these met-

rics is that they do not require an additional validation

dataset for the purpose of ﬁnding an optimal decision

threshold.

4 RESULTS

The quantitative results of the models trained on

healthy skin images for the task of skin lesion de-

tection are visualized in Figure 2. The ROC-Curves

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

538

(a) (b)

Figure 3: Visualization of (a) abnormal images with high anomaly scores, (b) abnormal images with low anomaly scores, (c)

normal images with high anomaly scores and (d) normal images with low anomaly scores based on the autoencoder.

show that both models achieve good results (Sim-

pleNet AUC 0.97, autoencoder AUC 0.93) and there-

fore are able to accurately detect skin anomalies. The

evaluation of the Precision-Recall-Curves yields sim-

ilar results (SimpleNet AUC 0.99, autoencoder AUC

0.97). In both cases, SimpleNet slightly outperforms

the autoencoder.

Furthermore, for qualitative analysis, we visual-

ize randomly drawn abnormal test images, their cor-

responding anomaly maps generated by both anomaly

detection methods and reconstruction images gener-

ated by the autoencoder in Figure 1. The anomaly

maps of the two compared models show major visual

differences. The anomaly maps generated by the au-

toencoder show ﬁne-grained details and are, to some

extent, very good at highlighting local skin patholo-

gies. However, some anomaly maps contain large ar-

eas of false positives, often corresponding to shading

or skin folds in the input image that have not been

correctly reconstructed by the autoencoder. In con-

trast, the anomaly maps generated by SimpleNet are

less detailed, but the region containing the anomalies

is in most cases roughly highlighted.

For further qualitative analysis we sorted all nor-

mal and abnormal test images by their anomaly score

in ascending order. Four abnormal images with the

lowest and highest anomaly score as well as four

normal images with the lowest and highest anomaly

scores are visualized in Figure 3 for the autoencoder

and in Figure 4 for SimpleNet, respectively. It can be

observed that both models assign low anomaly scores

to abnormal images containing scars or imprints of

clothing (see Figure 3 (b) and 4 (b)). This is reason-

able, because these images do not contain strong con-

trasts and therefore look similar to normal images. At

the same time, normal images with strong shading e.g.

over bony prominences tend to be assigned higher

anomaly scores (see Figure 3 (c) and 4 (c)). In con-

trast, normal images with low anomaly scores look

smooth without much variation (see Figure 3 (d) and

4 (d)). SimpleNet yields particularly high anomaly

scores for abnormal images containing skin lesions

that are bright red in colour and are a strong contrast

compared to the surrounding skin (see Figure 4 (a)).

In these cases, SimpleNet is also very good at local-

izing the abnormal region which can be observed in

the corresponding anomaly map. In contrast, the au-

toencoder assigns the highest anomaly score to abnor-

mal images with very dark shadows at the image bor-

ders (see Figure 3 (a)). It appears that the reconstruc-

tion error in these regions is so large that the actual

anomaly is barely detected. This can also be observed

in some examples in Figure 1.

5 DISCUSSION AND

CONCLUSION

The aim of this study was to explore how accurately

anomaly detection methods are able to detect and lo-

Semi-Supervised Anomaly Detection in Skin Lesion Images

539

(a) (b)

Figure 4: Visualization of (a) abnormal images with high anomaly scores, (b) abnormal images with low anomaly scores, (c)

normal images with high anomaly scores and (d) normal images with low anomaly scores based on SimpleNet.

calize different types of skin lesions after only being

presented images of healthy skin during training. To

answer this question a custom skin anomaly detection

dataset was created and two anomaly detection mod-

els (SimpleNet and autoencoder) were trained and

evaluated on this dataset. The results indicate that

both models are able to accurately distinguish ab-

normal images from normal images with SimpleNet

achieving an AUROC score of 97 % and the autoen-

coder a score of 93 %, respectively. Due to the ab-

sence of ground truth segmentation masks, quantita-

tive evaluation of the localization performance was

not possible. However, notable visual differences

were observed when comparing the anomaly maps

generated by each model.

Compared to SimpleNet, the autoencoder is bet-

ter at capturing ﬁne-grained anomaly details, due to

its reconstruction-based approach. In this approach,

each anomaly score is derived from the deviation be-

tween the original and reconstructed image pixel, al-

lowing ﬁner details to be preserved. In contrast, Sim-

pleNet calculates each anomaly score using a discrim-

inator neural network based on an image feature vec-

tor which describes the corresponding local neigh-

bourhood. As a result, details get lost during this pro-

cess.

However, the autoencoder showed high sensitiv-

ity to strong shading, frequently misclassifying it as

an abnormal region. This misclassiﬁcation occurs

when shading is poorly reconstructed, resulting in

high anomaly scores that, in some cases, exceed those

of actual abnormal regions. It is possible that im-

ages with poor lighting conditions were underrepre-

sented in the training dataset, contributing to this is-

sue. In this case it would be reasonable for the model

to classify shading as abnormal. However, as long as

the overall image is correctly classiﬁed as abnormal,

pixel-level misclassiﬁcations do not impact anomaly

detection metrics like AUROC. For this reason, the

localization accuracy should be investigated quanti-

tatively in future studies to explore if the image was

classiﬁed as abnormal for the right reasons.

In addition to challenges with reconstructing

strong shading, other issues arose, such as mis-

matches in skin tone between the original and recon-

structed images. In some cases, features such as skin

folds appeared in the reconstruction even though they

were absent in the original image. Against this back-

ground, it is important to note that the autoencoder

model used for inference was selected based on the

highest validation AUROC score. Thus, the empha-

sis was on optimizing the anomaly detection perfor-

mance, rather than achieving the best possible recon-

struction quality.

A limitation of our study lies in the small sam-

ple size and diversity of our dataset, which may re-

strict the generalization ability of our model. To

further explore semi-supervised anomaly detection

in dermatology, a larger medical dataset containing

healthy skin images would be required. This data

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

540

set should exhibit a high degree of variety regard-

ing factors such as age, skin tone, presence of body

parts, body hair, and different lighting conditions. Fu-

ture work could further explore anomaly-localization

performance, which would require additional ground-

truth masks of various skin anomaly types, created by

medical experts.

ACKNOWLEDGEMENTS

This research has received funding from the Ministry

of Science and Health of Rhineland-Palatinate, Ger-

many, and the Debeka Krankenversicherungsverein

a.G. through the Forschungskolleg Data2Health.

REFERENCES

Cai, Y., Zhang, W., Chen, H., and Cheng, K.-T. (2024). Me-

dianomaly: A comparative study of anomaly detection

in medical images. arXiv preprint arXiv:2404.04518.

Chan, S., Reddy, V., Myers, B., Thibodeaux, Q., Brown-

stone, N., and Liao, W. (2020). Machine learning in

dermatology: current applications, opportunities, and

limitations. Dermatology and therapy, 10:365–386.

Chandola, V., Banerjee, A., and Kumar, V. (2009).

Anomaly detection: A survey. ACM computing sur-

veys (CSUR), 41(3):1–58.

Emu, I. A., Niloy, N. T., Karim, B. M. A., Chowdhury,

A., Johora, F. T., Hasan, M., Mittra, T., Rashid, M.

R. A., Jabid, T., Islam, M., et al. (2024). ArsenicSkin-

ImageBD: A comprehensive image dataset to classify

affected and healthy skin of arsenic-affected people.

Data in Brief, 52:110016.

Gonzalez-Jimenez, A., Lionetti, S., Pouly, M., and

Navarini, A. A. (2023). Sano: Score-based diffusion

model for anomaly localization in dermatology. In

Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition, pages 2988–

2994.

Grignafﬁni, F., Troiano, M., Barbuto, F., Simeoni, P.,

Mangini, F., D’Andrea, G., Piazzo, L., Cantisani, C.,

Musolff, N., Ricciuti, C., et al. (2023). Anomaly

detection for skin lesion images using convolutional

neural network and injection of handcrafted features:

a method that bypasses the preprocessing of dermo-

scopic images. Algorithms, 16(10):466.

Liu, Z., Zhou, Y., Xu, Y., and Wang, Z. (2023). Simplenet:

A simple network for image anomaly detection and

localization. In Proceedings of the IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition,

pages 20402–20411.

Lu, Y. and Xu, P. (2018). Anomaly detection for skin

disease images using variational autoencoder. arXiv

preprint arXiv:1807.01349.

Shen, H., Chen, J., Wang, R., and Zhang, J. (2020). Coun-

terfeit anomaly using generative adversarial network

for anomaly detection. IEEE Access, 8:133051–

133062.

Sun, X., Yang, J., Sun, M., and Wang, K. (2016). A

benchmark for automatic visual classiﬁcation of clini-

cal skin disease images. In Computer Vision–ECCV

2016: 14th European Conference, Amsterdam, The

Netherlands, October 11-14, 2016, Proceedings, Part

VI 14, pages 206–222. Springer.

Tschuchnig, M. E. and Gadermayr, M. (2022). Anomaly

detection in medical imaging-a mini review. In Data

Science–Analytics and Applications: Proceedings

of the 4th International Data Science Conference–

iDSC2021, pages 33–38. Springer.

Zhang, H., Guo, W., Zhang, S., Lu, H., and Zhao, X. (2022).

Unsupervised deep anomaly detection for medical im-

ages using an improved adversarial autoencoder. Jour-

nal of Digital Imaging, 35(2):153–161.

Semi-Supervised Anomaly Detection in Skin Lesion Images

541