Channel-wise Aggregation with Self-correction Mechanism for

Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging

Mohamed Abdel-Nasser

1,3 a

, Adel Saleh

and Domenec Puig

1 b

Computer Engineering and Mathematics Department, University Rovira i Virgili, Tarragona, Spain

Gaist Solutions Ltd., U.K.

Electrical Engineering Department, Aswan University, Aswan, Egypt

Keywords:

Computational Pathology, Nuclei Segmentation, Whole Slide Imaging, Deep Learning.

Abstract:

In the ﬁeld of computational pathology, there is an essential need for accurate nuclei segmentation methods for

performing different studies, such as cancer grading and cancer subtype classiﬁcation. The ambiguous bound-

ary between different cell nuclei and the other objects that have a similar appearance beside the overlapping

and clumped nuclei may yield noise in the ground truth masks. To improve the segmentation results of cell

nuclei in histopathological images, in this paper, we propose a new technique for aggregating the channel maps

of semantic segmentation models. This technique is integrated with a self-correction learning mechanism that

can handle noisy ground truth. We show that the proposed nuclei segmentation method gives promising results

with images of different organs (e.g., breast, bladder, and colon)collected from medical centers that use de-

vices of different manufacturers and stains. Our method reaches the new state-of-the-art. Mainly, we achieve

the AJI score of 0.735 on the Multi-Organ Nuclei Segmentation benchmark, which outperforms the previous

closest approaches.

1 INTRODUCTION

Currently, digital pathology has an essential role in

clinics and laboratories. Thousands of tissue biop-

sies are taken from cancer patients yearly, and in turn,

the whole-slide imaging (WSI) technique permits the

acquisition of high-resolution images of slides. Cell

segmentation indicates the segmentation of the cell

nuclei. Notably, there is a necessity for precise nuclei

segmentation techniques in the context of computa-

tional pathology for facilitating the extraction of de-

scriptors for the morphometrics of cell nuclei. Several

descriptors, such as cell nuclei shape and number of

cell nuclei in WSI images, can be used to conduct var-

ious studies such as the determination of cancer types,

cancer grading, and prognosis (Moen et al., 2019).

Indeed, there is a diverse tissue type, variations in

staining and cell type leads to different visual char-

acteristics of WSI images. This makes the segmenta-

tion of nuclei segmentation a challenging task. Nuclei

segmentation task necessitates a vast effort to manu-

ally create the pixel-wise annotations that can be used

for training machine learning techniques. There are

nuclei segmentation toolboxes available in Cell Pro-

https://orcid.org/0000-0002-1074-2441

https://orcid.org/0000-0002-0562-4205

ﬁler (Carpenter et al., 2006) and ImageJ-Fiji (Schin-

delin et al., 2012). However, the visual characteris-

tics of WSI images makes it very difﬁcult to develop

traditional image processing based segmentation al-

gorithms that give acceptable nuclei segmentation re-

sults with WSI images taken from several cancer pa-

tients and collected at different medical centers for

various organs, such as breast, colon, lung, and stom-

ach (Niazi et al., 2019).

Recently, a review and comprehensive compar-

ison are presented in (Vicar et al., 2019) for cell

segmentation methods for label-free contrast mi-

croscopy. These segmentation methods studied are

categorized as follows: 1) single-cell segmentation

methods, 2) foreground segmentation methods, such

as thresholding, feature-extraction, level-set, graph-

cut, machine learning-based, and 3) seed-point ex-

traction methods, namely Laplacian of Gaussians, ra-

dial symmetry and distance transform, iterative radial

voting, maximally stable extremal region. The study

concluded that the machine learning-based methods

give accurate segmentation results.

In the last years, deep learning models have

been employed for performing different segmentation

tasks in biology (Niazi et al., 2019). Naylor et al.

(Naylor et al., 2017) introduced a fully automated

method for cell nuclei segmenting in histopathology

466

Abdel-Nasser, M., Saleh, A. and Puig, D.

Channel-wise Aggregation with Self-correction Mechanism for Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging.

DOI: 10.5220/0009156604660473

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

466-473

ISBN: 978-989-758-402-2; ISSN: 2184-4321

images based on three semantic segmentation mod-

els: PangNet,a fully convolutional network (FCN)

and DeconvNet. They ensembled the three seman-

tic segmentation models and obtained an F1-score of

0.80. In (Al-Kofahi et al., 2018), a three-step cell nu-

clei segmentation approach is proposed: 1) the detec-

tion of the cells using a deep learning-based model

to obtain pixel probabilities for nuclei, cytoplasm,

as well as background, 2) the separation of touching

cells based on blob detection and shape-based water-

shed techniques that can distinguish between the indi-

vidual nuclei from the nucleus prediction map and 3)

the segmentation of the nucleus and cytoplasm). With

four different datasets, they obtained an accuracy of

0.84.

In (Qu et al., 2019), a weakly supervised

deep nuclei segmentation using points annotation in

histopathology images is proposed. In this study, the

original WSI images and the shape prior of nuclei are

employed to obtain two types of coarse labels from

the points annotation using the Voronoi diagram and

the k-means clustering algorithm. These coarse la-

bels are used to train a deep learning model, and then

the dense conditional random ﬁeld is utilized in the

loss function to ﬁne-tune the trained model. With

the multi-organ WSI dataset, a dice score of 0.73 is

achieved. In (Mahmood et al., 2019), a conditional

generative adversarial network (cGAN) model is pro-

posed for nuclei segmentation. A large dataset of syn-

thetic WSI images with perfect nuclei segmentation

labels is generated using an unpaired GAN model.

Both synthetic and real data with spectral normaliza-

tion and gradient penalty for nuclei segmentation are

used to train the cGAN model.

In (Zhou et al., 2019), a deep learning-based

model called contour-aware informative aggregation

network (CIA-Net) with a multilevel information ag-

gregation module between two task-speciﬁc decoders.

Instead of using independent decoders, this model ex-

ploits bi-directionally aggregating task-speciﬁc fea-

tures to merge the advantages of spatial and texture

dependencies between nuclei and contour. Besides, a

smooth truncated loss that modulates losses is utilized

to mitigate the perturbation from outliers. As a result,

the CIA-Net model is almost built using informative

samples, and so its generalization capability could be

enhanced (i.e., with multi-organ multi-center nuclei

segmentation tasks). With the 2018 MICCAI chal-

lenge of the multi-organ nuclei segmentation dataset,

they produced a Jaccard score of 0.63.

In addition, in (Wang et al., 2019), a multi-path di-

lated residual network is proposed for nuclei segmen-

tation and detection. The network includes the fol-

lowing: 1) multi-scale feature extraction based on D-

Figure 1: WSI images.

ResNet and feature pyramid network (FPN), 2) can-

didate region network, and 3) a ﬁnal network for de-

tection and segmentation. The detection and segmen-

tation network involves three parts: segmentation, re-

gression, and classiﬁcation sub-networks. With the

MonuSeg dataset, an aggregated Jaccard index (AJI)

of 0.46 is obtained.

Although methods above achieved promising cell

nuclei segmentation results, the ambiguous bound-

ary between different cell nuclei and the other ob-

jects that have a similar appearance beside the over-

lapping and clumped nuclei may yield noise in the

ground truth masks (see Fig. 1). To cope with these

issues, in this paper, we propose a new technique for

aggregating the channel maps of semantic segmenta-

tion models. This technique is integrated with a self-

correction learning mechanism that can handle noisy

ground truth. Notably, we do not claim any novelty

of the self-correction mechanism or the channel-wise

aggregation mechanism, but only the superiority of

the performance of the proposed method with the nu-

clei segmentation task.

The rest of this paper is organized as follows. Sec-

tion 2 presents the proposed methods. Section 3 pro-

vides the results and discussion. Section 4 concludes

the paper and gives some points of future work.

2 METHODOLOGY

The ambiguous boundary between different cell nu-

clei and the other objects that have a similar appear-

ance beside the overlapping and clumped nuclei may

yield noise in the ground truth masks. To improve

the segmentation results of cell nuclei in histopatho-

logical images, we propose a new technique for ag-

Channel-wise Aggregation with Self-correction Mechanism for Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging

467

gregating the channel maps of semantic segmenta-

tion models. This technique is integrated with a self-

correction learning mechanism that can handle noisy

ground truth.

2.1 Channel-wise Aggregation

Mechanism

The aggregation operators reduces a set of numbers

, x

, . . . , x

) into a single representative number y.

This operation can be expressed as follows (Lucca

et al., 2017):

y = A(x

, x

, . . . , x

) (1)

here A : [0, 1]

×[0, 1] → [0, 1] is said to be aggregation

function Iff it has the following conditions:

• Identity in the Case of Unary: A(x) = x

• Boundary Conditions: A(0

, ..., 0

) = 0 and

agg(1

, ..., 1

) = 1

• Non Decreasing: A(x

, ..., x

) ≤ A(y

, ..., y

)

when (x

, ..., x

) ≤ (y

, ..., y

)

For the sake of simplicity, in this paper, we use the

maximum operator (max{x

, . . . , x

}). In the decoder

part of the segmentation model, we apply a element-

wise max. aggregation function on the channel maps.

If we have channel maps {C1, C2, . . . , C

}, the aggre-

gation function will produce one channel map (C

gg).

Then, we concatenate the C

gg with the original fea-

ture maps and feed all together to the next layers as

folows {C

gg, C1, C2, . . . , C

2.2 Self Correction Training

Mechanism

The self-correction training strategy (Li et al., 2019)

can be employed to aggregate a segmentation model

and labels, which can enhance the performance of

the segmentation model and the ground-truth labels

in an iterative manner. The improvement that can be

achieved by the self-correction training strategy de-

pends on the initial segmentation results of the basic

segmentation model. It is important to note that if the

initial segmentation results are not accurate, they may

worsen the self-correction training strategy. Thus,

the self-correction strategy should be begun after the

training loss begins to take a ﬂat shape.

After we get good segmentation results with

the basic segmentation model, a cyclically learning

scheduler with warm restarts is used. Here, a cosine

annealing learning rate scheduler with cyclical restart

(Loshchilov and Hutter, 2016) that can be mathemat-

ically formulated as follows:

φ = φ

min

(φ

max

− φ

min

)



1 + cos



res



(2)

where EP is the number of epochs in each cycle, EP

res

indicates the number of epochs gone since the previ-

ous restart, φ

max

is the initial learning rate, and φ

min

the ﬁnal learning rate.

After each cycle of the self-correction mecha-

nism, we obtain a set of weights (models) θ =



, . . . ,



, and the corresponding predicted la-

bels Y =

{

ˆy

, ˆy

, . . . , ˆy

}

, where T is the number of

training cycles. After each training cycle, the current

model weights

θ are aggregated with the weights of

previous cycle

t−1

in order to obtain new weights

as follows:

t + 1

t−1

t + 1

θ (3)

Similarly, the the ground-truth labels are be aggre-

gated as follows:

ˆy

t + 1

ˆy

t−1

t + 1

ˆy (4)

where t refers to the number of the current cycle (0 ≤

t ≤ T ), and ˆy is the generated pseudo-labels (pseudo-

masks) with the model

2.3 Basic Framework for Conducting

Cell Nuclei Segmentation

The self-correction training used in his paper utilizes

the A-CE2P model (Ruan et al., 2019) as the basic

framework for conducting cell nuclei segmentation.

As shown in Fig. 2, the CE2P model comprises three

branches: segmentation branch (top), fusion branch

(middle), and edge branch (bottom). The operation

of A-CE2P can be mathematically formulated as fol-

lows:

E = α

edge

+ α

parsing

+ α

consistent

(5)

where α

, α

and α

are hyper-parameters to control

the contribution among these three losses. The CE2P

model is jointly trained in an end-to-end manner by

minimizing E .

For an input WSI image img, assume that the cell

nuclei ground truth label is ˆy

and the predicted mask

is y

, where cn refers to the number of pixels for class

i. The pixel-level supervised nuclei segmentation task

can be expressed using the cross-entropy loss as fol-

lows:

= −

∑

ˆy

log p (y

) (6)

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

468

conv1

conv2

conv3

conv4

conv5

Context

Encoding

UpsampleConcat

Decoder Module

UpsampleConcat

Decoder Module

UpsampleConcat

Decoder Module

Loss

Edge

Mask

Loss

Segmentation Loss

EdgeLoss

ConsistencyConstrain

Decoder Module

conv

max channel

concat

Upsampling

conv

max channel

concat

Upsampling

* * *

Figure 2: Self correction training mechanism.

where N is the number of pixels, K is the num-

ber of classes. To let the model facilitate the

mean intersection-over-union (mIoU) directly, the the

cross-entropy loss and the mIoU loss are combined as

follows:

cells

= E

+ L

mIoU

(7)

To preserve the consistency between the predicted nu-

clei segmentation masks and the boundary prediction,

the following constraint term is exploited:

consistent

|N|

∑

n∈N

pos



edg

− edg



(8)

In this expression, N is the number of positive edge

pixels, edg

refers to the edge maps produced by the

edge branch and

edg

is the edge maps produced by

the nuclei segmentation branch y

. When computing

the loss, we should avoid that non-edge pixels domi-

nate the loss. To do so, the non-edge pixels are sup-

pressed, and only positive edge pixels cn ∈ N

pos

are

allowed to contribute to the consistency term. Here,

the ResNet-101 (He et al., 2016) is used as the back-

bone of the feature extractor with the ImageNet pre-

trained weights.

3 RESULTS AND DISCUSSION

3.1 Evaluation Metrics

Aggregated Jaccard Index (AJI) is proposed in (Ku-

mar et al., 2017) assessing the performance of nuclei

segmentation methods. If AJI equals 1, it means that

we obtain perfect nuclei segmentation results. AJI is

an modiﬁed version of the Jaccard index that divides

the aggregated intersection cardinality by the aggre-

gated union cardinality in the ground truth and seg-

mented masks. AJI can be expressed as follows:

AJI =

∑

i=1



∩ NP

∗

(i)



∑

i=1



∪ NP

∗

(i)



∑

k∈Ind

(9)

In this expression, GT =

i=1,2..K

is the ground-

truth of the nuclei pixels, NP =

j=1,2...L

are the

prediction nuclei segmentation results, NP

∗

(i) is the

connected component from the prediction result that

maximize the Jaccard index, and Ind is the list of in-

dices of pixels that do not belong to any component

in the GT.

Channel-wise Aggregation with Self-correction Mechanism for Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging

469

The F1-score is the harmonic mean between pre-

cision and recall, which an be formulated as follows:

F1 =

2.T P

2.T P + FP + FN

(10)

where TP, FP, FN are the true positive, false pos-

itive, and false negative rates, respectively. The true

negative (TP) rate is deﬁned as T N = GT ∩NP, which

is the area not belonging to any of the two masks GT

and NP.

3.2 Experimental Results and

Discussion

The dataset used in this study has been obtained from

(Kumar et al., 2017). This dataset includes 30 WSI

images with annotations from 7 organs (breast, kid-

ney, colon, stomach, prostate, liver, and bladder) col-

lected at different medical centers. The size of each

image is 1000 × 1000.

The test data includes one image from every organ

that was not exposed to the network. The rest of im-

ages is used for training. Every image in the training

and testing data was scaled to 1024x1024 and divided

into four non-overlapping patches of size 512x512.

Further random cropping of 512x512 from every im-

age was applied, as well. The overall training data

had 4906 of 512x512 patches. A batch size of one im-

age was used due the limitation in resources, namely

the GPU memory. Proposed model was trained for

50 epoch. The stochastic gradient descent (SGD) was

used as an optimizer with an initial learning rate as

1e−1, momentum as 0.99 and weight decay as1e−8.

A Titan X GPU was used to run the experiments Table

1 shows the results of the proposed method and ones

of ﬁve state-of-the-art semantic segmentation based

on deep learning models nuclei segmentation. The

ﬁve models are Fully Convolutional Network (FCN),

U-Net, Mask R-CNN, and conditional GAN (cGAN).

As shown the proposed method achieves an F1-score

of 0.876 and AJI score of 0.735. These results are bet-

ter than the ones of the previous approach (Mahmood

et al., 2019). We also show that the addition of the

channel wise aggregation improves the performance

of the baseline framework (self attention mechanism

with CE2P).

Figure 3 shows the segmentation results of the

proposed method with different organs: breast, kid-

ney, liver, prostate, bladder, colon, and stomach. As

shown, our method produces good segmentation re-

sults with bladder and stomach histopathological im-

ages with AJI scores of 0.85 and 0.83, respectively.

The proposed method gives a segmentation results

lower than 0.67 with the liver image because of the

apparent overlap between several cell nuclei.

Table 1: Comparison between the proposed model and the

related methods: FCN, U-Net, Mask R-CNN, and cGAN.

Method F1−score AJI

FCN (Long et al., 2015) 0.35 0.35

U-Net (Ronneberger et al., 2015) 0.41 0.41

Mask R-CNN (He et al., 2017) 0.50 0.50

cGAN (Mahmood et al., 2019) 0.87 0.72

Baseline 0.88 0.73

Proposed 0.89 0.74

Figure 4 shows a comparison between the number

of cell nuclei in the predicted masks and the corre-

sponding ground truth. As shown, the number of cell

nuclei obtained by the proposed method is a bit higher

than the ones of the ground-truth.

The proposed model gives promising segmenta-

tion results when we have noisy ground truth masks

because of the ambiguous boundary between differ-

ent cell nuclei and the other objects that have a simi-

lar appearance beside the overlapping and clumped.

If the cell nuclei ground-truth is almost clean, the

channel-wise aggregation, label reﬁnement, and the

self-correction training mechanisms can be seen as an

ensembling of clones of the basic segmentation model

(i.e, CE2P), which would improve the cell nuclei seg-

mentation results and produce a generalized model

that can be used with images of different organs ac-

quired at different medical centers.

4 CONCLUSION

In this paper, we propose a new technique for ag-

gregating the channel maps of semantic segmenta-

tion models in order to improve the segmentation re-

sults of cell nuclei in histopathological images. This

technique is integrated with a self-correction learning

mechanism that can handle noisy ground truth. We

show that the proposed nuclei segmentation method

gives promising results with images of different or-

gans (e.g., breast, bladder, and colon)collected from

medical centres that use devices of different manu-

facturers and stains. Our method achieves the new

state-of-the-art. Particularly, we achieve the AJI

score of 0.735 on the Multi-Organ Nuclei Segmen-

tation benchmark, which outperforms the previous

closest approaches. In the future work, we will ex-

plore the use of different aggregation functions to

improve the segmenting cell nuclei results. We will

also use the proposed segmentation model to segment

breast masses in other modalities, such as thermogra-

phy (Abdel-Nasser et al., 2016a; Abdel-Nasser et al.,

2016b).

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

470

Organ: Breast

Clinic:ChristianaHealthcare

AJI=0.72

Ground truth                                      Segmentation

Organ: Kidney

Clinic:University ofPittsburgh

AJI=0.70

Organ: Liver

Clinic:Princess Margaret Hospital

AJI=0.67

Organ: Prostate

Clinic:Roswell Park

AJI=0.72

Organ: Bladder

Clinic:Memorial Sloan

AJI=0.85

Organ: Colon

Clinic:NorthCarolina

AJI=0.73

Organ: Stomach

Clinic:HealthNetwork

AJI=0.83

Figure 3: Segmentation results of the proposed model with different organs: breast, kidney, liver, prostate, bladder, colon, and

stomach.

Channel-wise Aggregation with Self-correction Mechanism for Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging

471

0 5 10 15 20 25 30

Patch number

100

150

200

250

300

350

Number of Cells

Ground Truth Proposed Model

Figure 4: Number of cell nuclei in the predicted masks and the corresponding ground truth.

ACKNOWLEDGEMENTS

This research was partly supported by the Spanish

Govern-ment through project DPI2016-77415-R.

REFERENCES

Abdel-Nasser, M., Moreno, A., and Puig, D. (2016a). Tem-

poral mammogram image registration using optimized

curvilinear coordinates. Computer methods and pro-

grams in biomedicine, 127:1–14.

Abdel-Nasser, M., Saleh, A., Moreno, A., and Puig, D.

(2016b). Automatic nipple detection in breast ther-

mograms. Expert Systems with Applications, 64:365–

374.

Al-Kofahi, Y., Zaltsman, A., Graves, R., Marshall, W., and

Rusu, M. (2018). A deep learning-based algorithm

for 2-d cell segmentation in microscopy images. BMC

bioinformatics, 19(1):365.

Carpenter, A. E., Jones, T. R., Lamprecht, M. R., Clarke, C.,

Kang, I. H., Friman, O., Guertin, D. A., Chang, J. H.,

Lindquist, R. A., Moffat, J., et al. (2006). Cellproﬁler:

image analysis software for identifying and quantify-

ing cell phenotypes. Genome biology, 7(10):R100.

He, K., Gkioxari, G., Doll

ar, P., and Girshick, R. (2017).

Mask r-cnn. In Proceedings of the IEEE international

conference on computer vision, pages 2961–2969.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane,

A., and Sethi, A. (2017). A dataset and a technique for

generalized nuclear segmentation for computational

pathology. IEEE transactions on medical imaging,

36(7):1550–1560.

Li, P., Xu, Y., Wei, Y., and Yang, Y. (2019). Self-correction

for human parsing. arXiv preprint arXiv:1910.09777.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 3431–3440.

Loshchilov, I. and Hutter, F. (2016). Sgdr: Stochastic

gradient descent with warm restarts. arXiv preprint

arXiv:1608.03983.

Lucca, G., Sanz, J. A., Dimuro, G. P., Bedregal, B., Asi-

ain, M. J., Elkano, M., and Bustince, H. (2017).

Cc-integrals: Choquet-like copula-based aggregation

functions and its application in fuzzy rule-based

classiﬁcation systems. Knowledge-Based Systems,

119:32–43.

Mahmood, F., Borders, D., Chen, R., McKay, G. N., Sal-

imian, K. J., Baras, A., and Durr, N. J. (2019). Deep

adversarial training for multi-organ nuclei segmenta-

tion in histopathology images. IEEE transactions on

medical imaging.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

472

Moen, E., Bannon, D., Kudo, T., Graf, W., Covert, M., and

Van Valen, D. (2019). Deep learning for cellular im-

age analysis. Nature methods, page 1.

Naylor, P., La

e, M., Reyal, F., and Walter, T. (2017).

Nuclei segmentation in histopathology images using

deep neural networks. In 2017 IEEE 14th Inter-

national Symposium on Biomedical Imaging (ISBI

2017), pages 933–936. IEEE.

Niazi, M. K. K., Parwani, A. V., and Gurcan, M. N.

(2019). Digital pathology and artiﬁcial intelligence.

The Lancet Oncology, 20(5):e253–e261.

Qu, H., Wu, P., Huang, Q., Yi, J., Riedlinger, G. M., De,

S., and Metaxas, D. N. (2019). Weakly supervised

deep nuclei segmentation using points annotation in

histopathology images. In International Conference

on Medical Imaging with Deep Learning, pages 390–

400.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Ruan, T., Liu, T., Huang, Z., Wei, Y., Wei, S., and Zhao,

Y. (2019). Devil in the details: Towards accurate sin-

gle and multiple human parsing. In Proceedings of

the AAAI Conference on Artiﬁcial Intelligence, vol-

ume 33, pages 4814–4821.

Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V.,

Longair, M., Pietzsch, T., Preibisch, S., Rueden, C.,

Saalfeld, S., Schmid, B., et al. (2012). Fiji: an open-

source platform for biological-image analysis. Nature

methods, 9(7):676.

Vicar, T., Balvan, J., Jaros, J., Jug, F., Kolar, R., Masarik,

M., and Gumulec, J. (2019). Cell segmentation meth-

ods for label-free contrast microscopy: review and

comprehensive comparison. BMC bioinformatics,

20(1):360.

Wang, E. K., Zhang, X., Pan, L., Cheng, C.,

Dimitrakopoulou-Strauss, A., Li, Y., and Zhe, N.

(2019). Multi-path dilated residual network for nuclei

segmentation and detection. Cells, 8(5):499.

Zhou, Y., Onder, O. F., Dou, Q., Tsougenis, E., Chen, H.,

and Heng, P.-A. (2019). Cia-net: Robust nuclei in-

stance segmentation with contour-aware information

aggregation. In International Conference on Informa-

tion Processing in Medical Imaging, pages 682–693.

Springer.

Channel-wise Aggregation with Self-correction Mechanism for Multi-center Multi-Organ Nuclei Segmentation in Whole Slide Imaging

473