Ensemble Clustering for Histopathological Images Segmentation using

Convolutional Autoencoders

Ilias Rmouque

, Maxime Devanne

2 a

, Jonathan Weber

2 b

Germain Forestier

2 c

and C

edric Wemmert

1 d

ICube, University of Strasbourg, France

IRIMAS, University of Haute-Alsace, France

Keywords:

Digital Pathology, Deep Learning, Autoencoders, Histopathology, Ensemble Learning, Segmentation.

Abstract:

Unsupervised deep learning using autoencoders has shown excellent results in image analysis and computer

vision. However, only few studies have been presented in the ﬁeld of digital pathology, where proper labelling

of the objects of interest is a particularly costly and difﬁcult task. Thus, having a ﬁrst fully unsupervised

segmentation could greatly help in the analysis process of such images. In this paper, many architectures of

convolutional autoencoders have been compared to study the inﬂuence of three main hyperparameters:

(1) number of convolutional layers, (2) number of convolutions in each layer and (3) size of the latent space.

Different clustering algorithms are also compared and we propose a new way to obtain more precise results

by applying ensemble clustering techniques which consists in combining multiple clustering results.

1 INTRODUCTION

Pathology is essential for the diagnosis evaluation

and understanding of many underlying biological and

physiological mechanisms. It is usually a visual eval-

uation by pathologists of a tissue sample using a mi-

croscope to identify its structural properties. Cur-

rently, the visual evaluation of microscopic specimens

is largely an unassisted process, and the pathologist’s

accuracy is established through extensive training,

comparative analysis, peer quality control and per-

sonal experience. However, this ﬁeld has undergone

several technological revolutions in recent years with

the advent of virtual microscopy (conversion of glass

slides into high-resolution images called Whole Slide

Images - WSI), often referred to as ”digital pathol-

ogy”. Thus, major efforts have been made to design

image analysis tools, for example to identify basic

biological structures (stroma, immune cells, tumour,

etc.), in order to make it easier for doctors to (semi-

)automate the interpretation of slides. Meanwhile, au-

tomatic image analysis algorithms have recently made

extraordinary progress, particularly with the advent

https://orcid.org/0000-0002-1458-3855

https://orcid.org/0000-0002-3694-4703

https://orcid.org/0000-0002-4960-7554

https://orcid.org/0000-0002-4360-4918

of the deep learning methods introduced by Lecun et

al. (LeCun et al., 2015). Indeed, the performances of

these methods have exploded in recent years, allow-

ing the detection, classiﬁcation and segmentation of

objects of interest in images and particularly in med-

ical images with high precision (Bukala et al., 2020;

Ando and Hotta, 2021). But most of these approaches

operate in supervised mode, i.e. they require many ex-

amples in order to provide an effective model. How-

ever, obtaining quality annotations on histopatholog-

ical images remains very costly. For example, in the

ﬁeld of colorectal cancer WSI segmentation, Qaiser

et al. proposed a method based on persistent ho-

mology to classify tumour and non-tumour patches

from Hematoxylin & Eosin stained histology images

(Qaiser et al., 2016). To train their system, more than

18,000 annotated patches were needed. At the same

time, unsupervised approaches have shown their in-

terest in many applications for image analysis, such as

remote sensing (Liang et al., 2018; Mei et al., 2019).

Recently, they have also been applied to histopatho-

logical WSI analysis for cells segmentation (Junior

and Backes, 2021) or regions of interest classiﬁcation

(Figueira et al., 2020). In particular for cancer, the au-

thors in (Yamamoto et al., 2019) describe an unsuper-

vised approach for extracting interesting information

from WSI that obtains better accuracy than human for

prognostic prediction of prostate cancer recurrence.

Rmouque, I., Devanne, M., Weber, J., Forestier, G. and Wemmert, C.

Ensemble Clustering for Histopathological Images Segmentation using Convolutional Autoencoders.

DOI: 10.5220/0010835300003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages

933-940

ISBN: 978-989-758-555-5; ISSN: 2184-4321

933

In this paper, we are interested in automatic seg-

mentation in order to quickly extract regions of in-

terest (tumours for example) to make a more precise

analysis of these areas only. However, only few ap-

proaches on fully unsupervised segmentation of WSI

have been proposed. The ﬁrst attempt to segment

regions of interest from WSI without any prior in-

formation or examples has been performed in (Khan

et al., 2013). The authors highlight tissue morphol-

ogy in breast cancer histology images by calculating

a set of Gabor ﬁlters to discriminate different regions.

In (Fouad et al., 2017), the authors use mathemati-

cal morphology to extract ‘virtual-cells’ (e.g. super-

pixels), for which morphological and colour features

are calculated to then apply a consensus clustering al-

gorithm to identify the different tissues in the image.

More recently, a similar approach has been presented

in (Landini et al., 2019), adding a semi-supervised

self-training classiﬁer to the previous techniques that

enhances the results at the cost of partial supervision.

All these approaches propose to cluster the image

based on predeﬁned features. However, deep learn-

ing approaches, particularly via autoencoding archi-

tectures, make it possible to avoid manual deﬁnition

of features by calculating a condensed representation

of the image in a latent space by applying convolu-

tional ﬁlters. Unfortunately, as stated in (Raza and

Singh, 2018), most applications of autoencoders in

digital pathology were developed to perform cell seg-

mentation or nuclei detection (Xu et al., 2015; Hou

et al., 2019), or stain normalisation (Janowczyk et al.,

2017). Therefore, we propose here to study the po-

tential of these approaches for WSI tissue segmenta-

tion. The aim is to try to automatically identify clus-

ters corresponding to each type of tissue in the WSI

that could then be labelled by pathologists.

In this paper, we present a study on how convolu-

tional autoencoders perform on WSI segmentation by

comparing different approaches. First, different au-

toencoders architectures are compared to quantify the

importance of hyperparameters of interest (number of

convolutional layers, number of convolutions by layer

and size of the latent space). Then, a multi-resolution

approach using an ensemble clustering framework is

evaluated, to see if such ensemble techniques could

provide more accurate results.

2 METHODS

2.1 Convolutional Autoencoders

In this section, we explore of the use of convolu-

tional autoencoders to cluster WSI histopathological

images. For this, we present several experiments to

evaluate the importance of each hyperparameter.

As shown in Figure 1, a Convolutional AutoEn-

coder (CAE) is a deep convolutional neural network

composed of two parts: an encoder and a decoder.

The main purpose of the CAE is to minimise a loss

function L, evaluating the difference between the in-

put and the output of the CAE (usually Mean Squared

Error). Once this function is minimised, we can as-

sume that the encoder part builds up a suitable sum-

mary of the input data, in the latent space, as the de-

coder part is capable of reconstructing an accurate

copy of it from this encoded representation.

The encoder is ﬁrst constituted of the input layer

(having the size of the input image) which is con-

nected to N convolutional layers of diminishing size,

up to an information bottleneck of size Z, called the

latent space. The bottleneck is connected to a se-

ries of N convolutional layers of increasing size, un-

til reaching the size of the input. This second part

is called the decoder. Each convolution layer is com-

posed of C convolutions and is followed by three other

layers: a batch normalisation, an activation function

(ReLU) and a max pooling of size (2,2).

To perform the clustering, a trained CAE is used to

encode each patch of the whole image. Then, this en-

coded representation of the patch (in the latent space)

is given as the input of a clustering algorithm and a

cluster is assigned to the patch.

We decided to evaluate the inﬂuence of the three

hyperparameters N, Z and C. For each one, different

values were tested while ﬁxing the two others (N =

2, Z = 250, C = 10). To evaluate the quality of the

results, the Adjusted Rand Index (ARI) is calculated

to compare the obtained clustering to the annotations

of the expert. The Rand Index computes a similarity

measure between two clusterings by considering all

pairs of samples and counting pairs that are assigned

in the same or different clusters in the predicted and

true clusterings. The score is then normalised into the

ARI score by:

ARI =

(RI − Expected RI)

(max(RI) − Expected RI)

(1)

Values of the ARI are close to 0 for random la-

belling independently of the number of clusters and

samples, and exactly 1 when the clusterings are iden-

tical (up to a permutation).

Each CAE was trained over a set of 10,000 differ-

ent patches randomly selected. As the result of both

the clustering and the training of the CAE are non-

deterministic, due to a high sensitivity to the initial

conditions, 10 autoencoders were trained and the re-

sults averaged for each hyperparameter value.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

934

Figure 1: Architecture of a CAE with N = 2, C = 10 and Z = 50.

We also investigated the performance of several

clustering algorithms, i.e Kmeans, Agglomerative

clustering (AggCl), Gaussian mixture (GM) and also

the not too deep clustering method (N2D) exposed in

(McConville et al., 2020). A clustering performed

directly with the Kmeans algorithm on the raw data

(without any data reduction by the CAE) has been cal-

culated as a baseline to evaluate the beneﬁt of encod-

ing the data with the CAE.

2.2 Ensemble Clustering

As exposed in (Yamamoto et al., 2019), both micro-

structures and macro-structures give different infor-

mation. Pathologists also agree that identifying a sin-

gle cell is way more difﬁcult without its surrounding

context and they always look at the WSI at lower

magniﬁcation (to better capture the context) before

zooming in at high magniﬁcation. Furthermore, in

(Alsubaie et al., 2018) an example of multi-resolution

lung cancer adenocarcinoma classiﬁcation using deep

learning shows improvements in the overall accuracy.

Thus, we explored a way to improve the results by

using an ensemble of clustering methods, each focus-

ing on a different resolution. The objective is to merge

low level information (context) with high level infor-

mation (shape of the cells, etc.). For this, the con-

sensus method proposed in (Wemmert and Ganc¸arski,

2002) was used. This method is based on a the eval-

uation of the similarity between different clusterings

and the deﬁnition of corresponding clusters. Then, a

multi-view voting approach is computed to produce

a single result representing all clusterings. An exam-

ple of the architecture of the approach is depicted in

Figure 2.

We explored different conﬁgurations, but we only

present the two most representative which highlight

how the quality of the results can be improved by

using ensemble clustering. The ﬁrst conﬁguration,

multires

is composed of three clustering algorithms

(Kmeans) working on the latent space representation

of the image obtained by different CAE trained at dif-

ferent resolutions: 10× with 8 clusters, 5× with 6

clusters and 5× with 8 clusters. As the reconstructed

image from the autoencoder seems to focus more on

colour intensity than real structures, a second ensem-

ble conﬁguration has been tested. To add diversity

and to force the ﬁnal result to focus its attention more

on the structure of the objects, a clustering working on

a binary image (by thresholding the intensity of the

initial image) has been computed. Thus, the second

conﬁguration (E

struct

) is composed of three clustering

algorithms (Kmeans) with the following parameters:

5× on the binary image with 6 clusters, 5× on the

binary image with 8 clusters and 10× on the initial

RGB image with 6 clusters.

3 EXPERIMENTS AND RESULTS

Our study was performed on 8 WSI of Haematoxylin

Eosin Saffron (HES) stained tissue extracted from

a cohort of patients built within the scope of the

AiCOLO project (INSERM/Plan Cancer) studying

colon cancer. The images have been provided by

Georges Franc¸ois Leclerc Centre (Dijon, France) and

acquired from two different centres. An example is

given in Figure 3a. HES stain distinguishes cell nuclei

in purple, from extracellular matrix and cytoplasm in

pink.

All images have been acquired at 20× magni-

ﬁcation (corresponding to 0.5 µm/pixel) but stored

at several resolutions in a pyramidal format. The

size of each image is around 90, 000 × 50, 000 pix-

els. To train autoencoders, 10, 000 patches of size

128× 128 pixels were randomly extracted at 10× res-

olution from all images (and 5× for the ensemble

approach), as this seems to be the minimal amount

Ensemble Clustering for Histopathological Images Segmentation using Convolutional Autoencoders

935

Figure 2: Architecture of the ﬁrst ensemble conﬁguration E

multires

: two CAE trained at different resolutions (10× and 5×)

produce different latent representations that are clustered. The three resulting clusterings are then merged through the multi-

view voting algorithm proposed in (Wemmert and Ganc¸arski, 2002).

(a) Example of a WSI of colon tissue stained with HES

(magniﬁcation: ×20, size: 97,920× 55,040 pixels)

(b) Example of clustering with 8 clusters (orange, red and

blue clusters corresponding to tumour)

Figure 3: Example of (a) a raw WSI and (b) a clustering result of this image.

of information required by human expert to classify

the tissue. Meanwhile, sparse manual annotations

of the ﬁve classes of tissue, tumour, stroma, outer

layer mucosa (crypts of Lieberkuhn and connective

tissue), immune cells, and necrosis, and two classes

for background and artefacts (ink marks, etc.) have

been performed by pathologists on the images (using

Cytomine(Mar

ee et al., 2016)), to be able to evaluate

the relevance of the clustering.

3.1 Evaluation of All Hyperparameters

of the CAE

First, results obtained without using the latent space

representation (see Table 1) are worse than all those

obtained when clustering the encoded data. This con-

ﬁrms the interest of using a CAE for WSI clustering.

As shown in Figure 4a, it appears that the number of

convolutions in each layer of convolutions (hyperpa-

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

936

1 3 5 7 10 15

0.0 0.1 0.2 0.3 0.4

7 Clusters

C : Number of convolutions

ARI

1 3 5 7 10 15

0.0 0.1 0.2 0.3 0.4

8 Clusters

C : Number of convolutions

ARI

1 3 5 7 10 15

0.0 0.1 0.2 0.3

9 Clusters

C : Number of convolutions

ARI

1 3 5 7 10 15

0.0 0.1 0.2 0.3

10 Clusters

C : Number of convolutions

ARI

(a) Number of convolutions C in each layer of convolutions (N = 2, Z = 250)

1 2 3 4 5

0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

7 Clusters

N : Number of layers

ARI

1 2 3 4 5

0.30 0.35 0.40 0.45

8 Clusters

N : Number of layers

ARI

1 2 3 4 5

0.30 0.35 0.40 0.45

9 Clusters

N : Number of layers

ARI

1 2 3 4 5

0.30 0.35 0.40 0.45

10 Clusters

N : Number of layers

ARI

(b) Number of layers of convolutions N (Z = 250, C = 10)

Figure 4: Evaluation of the ARI for the two main hyperparameters of the convolutions of the CAE comparing Kmeans

clustering on 7, 8, 9 and 10 clusters.

Ensemble Clustering for Histopathological Images Segmentation using Convolutional Autoencoders

937

10 25 50 100 250 500 1000

0.0 0.1 0.2 0.3 0.4

7 Clusters

Z : latent space size

ARI

10 25 50 100 250 500 1000

0.0 0.1 0.2 0.3 0.4

8 Clusters

Z : latent space size

ARI

10 25 50 100 250 500 1000

0.0 0.1 0.2 0.3 0.4

9 Clusters

Z : latent space size

ARI

10 25 50 100 250 500 1000

0.0 0.1 0.2 0.3 0.4

10 Clusters

Z : latent space size

ARI

Figure 5: Evaluation of the ARI with different latent space sizes, comparing Kmeans clustering on 7, 8, 9 and 10 clusters.

rameter C) does not greatly affects the quality of the

autoencoder as only a apart from a slight narrowing of

the variability of the results. It’s quite easy to ﬁgure

out why: passed a certain number, additional convo-

lution brings to few complementary information. Fig-

ure 4b shows the evaluation of the ARI with different

number of convolution layers in the architecture. We

can notice an increase of the quality index up to 4 lay-

ers and then a brutal drop at 5. This indicates clearly

that too many convolutions (and poolings that down-

sample the information) reduce the information that

can further not be properly processed.

Nonetheless, as seen in Figure 5, the latent space

size Z, seems to greatly inﬂuence the pertinence of the

CAE. Indeed, the ARI clearly grows as there is more

space to encode the latent representation, as a more

precise information can be stored. Also, the more in-

formation is present in the latent representation, the

more classes can easily be differentiated. However,

it is also clear that a too large latent space will not

be able to summarise efﬁciently the information, and

thus, will not help the clustering algorithm to discrim-

inate the different tissues. Moreover, the larger the la-

tent space, the more memory and time are needed to

train the network.

3.2 Comparison of the CAE with the

Ensemble Approach

As seen in the previous experiment, the ARI tends to

give low scores because we only have very few an-

notations on each class of interest. So we decided to

compute a second evaluation criterion based on the

ability of the clustering to detect tumours areas in the

image, as it is the main class of interest in our project.

To associate the tumour class to a cluster, we calcu-

lated its tumour density (number of labelled tumour

pixels / number of total labelled pixels in the clus-

ter). All clusters having a density over 50% are kept

as ’tumour’, the others are labelled as ’not tumour’.

Thus, two evaluation criteria have been calculated on

the results and are presented in Table 1: the ARI as

in the previous experiment (see Eq.1) and the FScore

on the two-classes problem (tumour vs. not tumour)

(Van Rijsbergen, 1979).

3.3 Discussion

Classical methods applied on the latent space rep-

resentation of the CAE tend to show acceptable re-

sults. However, both ensemble clustering conﬁgura-

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

938

Table 1: Evaluation of the ARI and FScore of all clustering results obtained with the different methods.

Raw data Encoded data

Kmeans Kmeans AggCl GM N2D E

multires

struct

Image 1 0.39 0.89 0.48 0.96 0.38 0.94 0.27 0.73 0.43 0.97 0.47 0.96 0.42 0.88

Image 2 0.27 0.68 0.33 0.63 0.29 0.62 0.19 0.43 0.29 0.68 0.31 0.66 0.46 0.62

Image 3 0.25 0.76 0.39 0.87 0.35 0.85 0.22 0.76 0.31 0.88 0.37 0.87 0.45 0.91

Image 4 0.08 0.48 0.08 0.50 0.13 0.61 0.05 0.51 0.12 0.55 0.08 0.54 0.08 0.75

Image 5 0.11 0.65 0.11 0.64 0.10 0.60 0.10 0.62 0.11 0.65 0.12 0.65 0.17 0.72

Image 6 0.37 0.68 0.52 0.75 0.51 0.72 0.49 0.67 0.43 0.76 0.51 0.77 0.57 0.75

Image 7 0.28 0.68 0.35 0.73 0.33 0.75 0.14 0.61 0.37 0.74 0.41 0.76 0.36 0.84

Image 8 0.33 0.63 0.44 0.71 0.42 0.70 0.07 0.44 0.37 0.69 0.44 0.75 0.45 0.75

Mean 0.26 0.68 0.34 0.72 0.31 0.72 0.19 0.60 0.30 0.74 0.34 0.75 0.37 0.78

Stdv 0.11 0.12 0.16 0.14 0.14 0.12 0.14 0.13 0.13 0.13 0.16 0.13 0.16 0.09

tions seem to be more efﬁcient in ﬁnding coherent

clusters corresponding to the classes of interest de-

ﬁned by the pathologists.

Among all the exposed methods, E

struct

seems to

give the best results. It tends to conﬁrm the impor-

tance of the shape of the objects on histopathological

images. Furthermore, it shows that even if convolu-

tional autoencoders aim at automatically ﬁnding the

best features to encode images, they can also take ad-

vantage of pre-computed features for some speciﬁc

tasks.

4 CONCLUSION

In this paper, we compared different conﬁgurations

of convolutional autoencoders in the ﬁeld of unsuper-

vised learning for WSI histopathological image seg-

mentation. For this, different CAE architectures have

been compared to try to ﬁnd the best conﬁguration

and to study the inﬂuence of each hyperparameter.

Then, we proposed a new approach that uses ensem-

ble clustering technique to take advantage of multires-

olution information and structural features in the im-

age. This conﬁrms the importance of having diversity

in an ensemble learning framework and that working

at different resolutions at the same time can really im-

prove the quality of the results.

ACKNOWLEDGEMENT

This work was supported by the AiCOLO project

funded by INSERM/Plan Cancer (MIC 2019).

REFERENCES

Alsubaie, N., Shaban, M., Snead, D., Khurram, A., and

Rajpoot, N. (2018). A multi-resolution deep learn-

ing framework for lung adenocarcinoma growth pat-

tern classiﬁcation. In Annual Conference on Medi-

cal Image Understanding and Analysis, pages 3–11.

Springer.

Ando, T. and Hotta, K. (2021). Cell image segmentation by

feature random enhancement module. In Proceedings

of the 16th International Joint Conference on Com-

puter Vision, Imaging and Computer Graphics Theory

and Applications - Volume 4: VISAPP,, pages 520–

527. INSTICC, SciTePress.

Bukala, A., Cyganek, B., Koziarski, M., Kwolek, B., Ol-

borski, B., Antosz, Z., Swadzba, J., and Sitkowski,

P. (2020). Classiﬁcation of histopathological images

using scale-invariant feature transform. In Farinella,

G. M., Radeva, P., and Braz, J., editors, Proceedings

of the 15th International Joint Conference on Com-

puter Vision, Imaging and Computer Graphics The-

ory and Applications, VISIGRAPP 2020, Volume 5:

VISAPP, Valletta, Malta, February 27-29, 2020, pages

506–512. SCITEPRESS.

Figueira, G., Wang, Y., Sun, L., Zhou, H., and Zhang,

Q. (2020). Adversarial-based domain adapta-

tion networks for unsupervised tumour detection in

histopathology. In 2020 IEEE 17th International Sym-

posium on Biomedical Imaging (ISBI), pages 1284–

1288. IEEE.

Fouad, S., Randell, D., Galton, A., Mehanna, H., and Lan-

dini, G. (2017). Unsupervised morphological segmen-

tation of tissue compartments in histopathological im-

ages. PloS one, 12(11):e0188717.

Hou, L., Nguyen, V., Kanevsky, A. B., Samaras, D., Kurc,

T. M., Zhao, T., Gupta, R. R., Gao, Y., Chen, W.,

Foran, D., et al. (2019). Sparse autoencoder for

unsupervised nucleus detection and representation in

histopathology images. Pattern recognition, 86:188–

200.

Janowczyk, A., Basavanhally, A., and Madabhushi, A.

(2017). Stain normalization using sparse autoencoders

Ensemble Clustering for Histopathological Images Segmentation using Convolutional Autoencoders

939

(stanosa): application to digital pathology. Computer-

ized Medical Imaging and Graphics, 57:50–61.

Junior, J. D. D. and Backes, A. R. (2021). Unsuper-

vised segmentation of leukocytes images using parti-

cle swarm. In VISIGRAPP (4: VISAPP), pages 439–

446.

Khan, A. M., El-Daly, H., Simmons, E., and Rajpoot, N. M.

(2013). Hymap: A hybrid magnitude-phase approach

to unsupervised segmentation of tumor areas in breast

cancer histology images. Journal of Pathology Infor-

matics, 4.

Landini, G., Fouad, S., Randell, D., and Mehanna, H.

(2019). Epithelium and stroma segmentation using

multiscale superpixel clustering. Journal of Pathology

Informatics, 10.

LeCun, Y., Bengio, Y., and Hinton, G. E. (2015). Deep

learning. Nature, 521(7553):436–444.

Liang, P., Shi, W., and Zhang, X. (2018). Remote sens-

ing image classiﬁcation based on stacked denoising

autoencoder. Remote Sensing, 10(1):16.

Mar

ee, R., Rollus, L., St

evens, B., Hoyoux, R., Louppe,

G., Vandaele, R., Begon, J.-M., Kainz, P., Geurts, P.,

and Wehenkel, L. (2016). Collaborative analysis of

multi-gigapixel imaging data using Cytomine. Bioin-

formatics, 32(9):1395–1401.

McConville, R., Santos-Rodriguez, R., Piechocki, R. J., and

Craddock, I. (2020). N2d: (not too) deep clustering

via clustering the local manifold of an autoencoded

embedding.

Mei, S., Ji, J., Geng, Y., Zhang, Z., Li, X., and Du, Q.

(2019). Unsupervised spatial–spectral feature learn-

ing by 3d convolutional autoencoder for hyperspectral

classiﬁcation. IEEE Transactions on Geoscience and

Remote Sensing, 57(9):6808–6820.

Qaiser, T., Sirinukunwattana, K., Nakane, K., Tsang, Y.-W.,

Epstein, D., and Rajpoot, N. (2016). Persistent homol-

ogy for fast tumor segmentation in whole slide his-

tology images. Procedia Computer Science, 90:119–

124.

Raza, K. and Singh, N. K. (2018). A tour of unsuper-

vised deep learning for medical image analysis. arXiv

preprint arXiv:1812.07715.

Van Rijsbergen, C. (1979). Information retrieval: the-

ory and practice. In Proceedings of the Joint

IBM/University of Newcastle upon Tyne Seminar on

Data Base Systems, pages 1–14.

Wemmert, C. and Ganc¸arski, P. (2002). A multi-view vot-

ing method to combine unsupervised classiﬁcations.

In Artiﬁcial Intelligence and Applications, pages 447–

452.

Xu, J., Xiang, L., Liu, Q., Gilmore, H., Wu, J., Tang,

J., and Madabhushi, A. (2015). Stacked sparse au-

toencoder (ssae) for nuclei detection on breast cancer

histopathology images. IEEE transactions on medical

imaging, 35(1):119–130.

Yamamoto, Y., Tsuzuki, T., Akatsuka, J., Ueki, M.,

Morikawa, H., Numata, Y., Takahara, T., Tsuyuki, T.,

Tsutsumi, K., Nakazawa, R., et al. (2019). Automated

acquisition of explainable knowledge from unanno-

tated histopathology images. Nature communications,

10(1):1–9.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

940