Field of Interest Proposal for Augmented Mitotic Cell Count:

Comparison of Two Convolutional Networks

Marc Aubreville

, Christof A. Bertram

, Robert Klopﬂeisch

and Andreas Maier

Pattern Recognition Lab, Computer Sciences, Friedrich-Alexander-Universit

at Erlangen-N

urnberg, Germany

Institute of Veterinary Pathology, Freie Universit

at Berlin, Germany

Keywords:

Mitotic Figure, Cell Segmentation, Digital Histopathology, Tumor Grading.

Abstract:

Most tumor grading systems for human as for veterinary histopathology are based upon the absolute count of

mitotic ﬁgures in a certain reference area of a histology slide. Since time for prognostication is limited in a

diagnostic setting, the pathologist will oftentimes almost arbitrarily choose a certain ﬁeld of interest assumed

to have the highest mitotic activity. However, as mitotic ﬁgures are commonly very sparse on the slide and

often have a patchy distribution, this poses a sampling problem which is known to be able to inﬂuence the

tumor prognostication. On the other hand, automatic detection of mitotic ﬁgures can’t yet be considered

reliable enough for clinical application. In order to aid the work of the human expert and at the same time

reduce variance in tumor grading, it is beneﬁcial to assess the whole slide image (WSI) for the highest mitotic

activity and use this as a reference region for human counting.

For this task, we compare two methods for region of interest proposal, both based on convolutional neural

networks (CNN). For both approaches, the CNN performs a segmentation of the WSI to assess mitotic activity.

The ﬁrst method performs a segmentation of mitotic cells at the original image resolution, while the second

approach performs a segmentation operation at a signiﬁcantly reduced resolution, cutting down on processing

complexity.

We evaluate the approach using a dataset of 32 completely annotated whole slide images of canine mast cell

tumors, where 22 were used for training of the network and 10 for test. Our results indicate that, while

the overall correlation to the ground truth mitotic activity is considerably higher (0.936 vs. 0.829) for the

approach based upon the ﬁne resolution network, the ﬁeld of interest choices are only marginally better. Both

approaches propose ﬁelds of interest that contain a mitotic count in the upper quartile of respective slides.

1 INTRODUCTION

Mitotic ﬁgures, i.e. cells undergoing cell division,

are an important marker for tumor prognostication, as

their density within tissue on a histology slide is as-

sumed to be correlated with the proliferative rate of

the tumor (Elston and Ellis, 1991). Hence it is not

surprising, that detection of mitotic ﬁgures has been

the target of several object detection challenges in re-

cent time (Roux et al., 2013; Veta et al., 2015; Veta

et al., 2018). Detection of mitotic ﬁgures in digi-

tal whole slide images (WSI) is, however, not only

a time-consuming task (as WSIs typically have very

large image dimensions), but also a task presently

not solved with a clinical applicable accuracy. This

can be related to a number of factors: Firstly, the

very deﬁnition of mitotic ﬁgures in histology slides

is tricky, as their morphology is vaguely described

as being without a nuclear membrane (post prophase)

with hairy extensions of nuclear material around the

chromosomes (Van Diest et al., 1992). Depending

on factors such as inferior tissue quality often deriv-

ing from delayed tissue ﬁxation, it is not always pos-

sible to unambiguously differentiate mitotic ﬁgures

from mitotic-like structures such as pyknotic tumor

cells of overstained nuclei. This leads to a high intra-

observer variance (Boiesen et al., 2000) in grading of

cells between labs, schools and even individuals that

are likely to reﬂect in data sets of mitotic ﬁgures ap-

plied for current developments of algorithms. Sec-

ondly, histology slides are subject to staining in or-

der to make important details visible to the human

eye. This dying procedure is however also subject to

a number of inﬂuencing factors, including concentra-

tion and purity (Horobin, 1969) of coloring agents,

slice thickness, the dying protocol and the dyed tis-

Aubreville, M., Bertram, C., Klopﬂeisch, R. and Maier, A.

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks.

DOI: 10.5220/0007365700300037

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 30-37

ISBN: 978-989-758-353-7

sue itself. This leads to a signiﬁcant color variance in

hematoxylin and eosin stained tissue sections, which

poses a challenge to pattern recognition methods, es-

pecially when color nuances may be a determining

factor for cell classiﬁcation. Lastly, mitotis is a sparse

event in histology slides, which in turn leads to low

numbers of events across databases and one can hy-

pothesize, that not the complete biological variance

spread is represented in current mitosis data sets like

the Mitos (Roux et al., 2013) or TUPAC (Veta et al.,

2018) data set.

Since mitotic ﬁgures can not be assumed as being

evenly spread over the image, manual count within

the usual diagnostic area of 10 consecutive High-

Power-Fields (HPF, ﬁeld of view at magniﬁcation of

400×) leads to an inherent sampling problem, as

also assumed by Bonert and Tate (Bonert and Tate,

2017). It is thus strongly dependent on the actual re-

gion chosen intuitively by the pathologist, how many

mitotic ﬁgures will be present within that area. Most

grading schemes incorporate the mitotic count (MC,

number of mitotic ﬁgures within 10 HPF) into the

tumor grade, often using a direct thresholding ap-

proach. Especially for tumors with borderline MC

around these thresholds, the area selection thus leads

to a signiﬁcant additional variance in the process of

grading.

We assume that, in order to be clinical applicable,

one interesting methodological approach would not

be the direct recognition and fully automated count of

mitotic ﬁgures in slides, as commonly performed, but

rather the determination of a region of interest with

a high mitotic ﬁgure density, assuming that this is

also the region with highest proliferation. It is gener-

ally assumed that the region with highest proliferation

has the strongest prognostic value for tumor grading

(Martin et al., 1995; Baak et al., 2008; Edmondson

et al., 2014).

As such, the primary output of our approach will

be the mitotic density of a given WSI. In order to

do so, we depend on an intermediate mitotic ﬁg-

ure segmentation map, which will be predicted by a

deep convolutional network. In previous work, we

have shown that the U-Net network architecture by

Ronneberger et al. (Ronneberger et al., 2015) is a

very good candidate for this approach (Aubreville

et al., 2018b). This model, however, comes with a

quite high inherent complexity, and we wondered if

a smaller version of the same approach directly tar-

geting at a subsampled image map could yield similar

overall performance.

2 MATERIAL

We annotated 32 whole slide images of canine cu-

taneous mast cell tumor, dyed with standard hema-

toxylin and eosin stain. All specimen was taken for

routine tumor diagnostics, therefore no IRB approval

was needed for this study. All slides were digitized

using a linear scanner (Aperio ScanScope CS2, Le-

ica Biosystems, Nussloch, Germany) at a magniﬁca-

tion of 400×, resulting in a digital resolution of 0.25

microns per pixel. We used the open source soft-

ware solution — (Aubreville et al., 2018a) to attain

a complete annotation map of all 32 WSI. The anno-

tation process was performed in a partly computer-

aided procedure, where the software would suggest

partly overlapping segments of the whole slide im-

age to the expert to annotate mitotic ﬁgures. In this

process, we did not only annotate mitotic ﬁgures, but

also granulocytes and other interesting cell types. It

should be noted that also non-mitotic cells with simi-

lar appearance to mitotic ﬁgures were annotated with

a designated class assigned to them. A second expert

was asked to rate all cells blindly (i.e. not knowing the

assigned class by the ﬁrst expert). We only consider

mitotic ﬁgures where both experts agreed on it being

a mitosis for our data set, however, for hard negative

examples, also mitotic ﬁgures annotated from one ex-

pert only or the aforementioned mitosis-like cells will

be part of our training process. Following this deﬁ-

nition of mitotic ﬁgure, our data set includes a total

of 45,811 mitotic cells. To increase generalization,

the data set purposefully includes cell tumors of dif-

ferent sizes and tumor grade, and thus the MC varies

tremendously across cases.

3 METHODS

Mitosis detection is often considered an object de-

tection approach, where singular events on an image

have to be counted (Veta et al., 2015; Li et al., 2018;

Cires¸an et al., 2013, and others..). This is due to

the fact that mitotic events are often seen as a sin-

gular occurrence that can be described using a single

(x,y) tuple. This is also reﬂected in several data sets

such as MICCAI AMIDA 2013 (Veta et al., 2015),

ICPR MITOS-ATYPIA 2014 and TUPAC 16 (Veta

et al., 2018), which use this for annotation. Other

data sets, such as the Mitos 2012 data set (Roux et al.,

2013), provide segmentation information for mitotic

cells, which is, however, a tedious process. In gen-

eral, dataset creation for mitotic ﬁgure detection tasks,

is a labour-intensive task, which might be one of the

reasons for the limited data set size. To reduce the

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks

whole slide

image

Mitosis

detector

network

RGB to

grey

w,h

closing

threshold

constrained

arg max

center

of region

proposal

w,h

Otsu

thres.

patch

extraction

concatate

mitotic figure

map

mitotic activity estimation branch

valid mask generation

Figure 1: Overview of the proposed approach for mitotic count region proposal (Aubreville et al., 2018b). The upper path

will derive singular mitotic annotations, followed by a moving average (MA) ﬁlter. The lower path derives an activity map of

the image to exclude border regions of the image. This paper compares two mitosis detector CNNs, as further detailed in Fig.

2 and Fig. 3.

impact of this, we decided to use our own data set

of canine mast cell tumors for this work. Addition-

ally to an unprecedented size, our data set provides

us with complete annotations of whole-slide-images,

so border regions of the tumor as well as regions not

containing tumor tissue will be included and enable

an increased robustness of the approach.

3.1 Field of Interest Proposal

The goal of the algorithm is to suggest an area of the

size of 10 adjacent High Power Fields with the high-

est mitotic count. Following Meuten et al., we as-

sume this area to be a total of 2.37 mm

(Meuten et al.,

2016). We use an aspect ratio of 4 : 3 for this rectan-

gular selection.

As depicted in Fig. 1, our approach consists of

the generation of a map of mitotic ﬁgures on the WSI

M as well as a map of valid tissue V . For estimation

of the mitotic count we utilize a convolutional neural

network for generation of segmentation maps of mi-

totic ﬁgures. In order to retrieve the mitotic activity

in a certain area, a moving average operator is used.

3.1.1 Mitotic Activity Estimation

For estimation of mitotic activity, the image is di-

vided into overlapping (margin: 64 px) patches with

a size of 512 × 512 px. Due to the structure of the net-

work, also other sizes would be applicable, reducing

efforts for not covering the overlapping margins mul-

tiple times, but increasing memory footprint on the

graphics card. The prediction of the network is be-

ing concatenated to yield an overall map M of mitotic

ﬁgure activity.

3.1.2 Valid Mask Estimation

In order to exclude regions of the image that are partly

uncovered by specimen, we construct a binary mask

of tissue presence from the WSI at a low magniﬁca-

tion. The image is converted to grey-scale, then a

binary threshold is performed using Otsu’s adaptive

method (Otsu, 1979). A closing operator is applied

to reduce thin interruptions of the tissue map, and ﬁ-

nally a moving average ﬁlter according to the size of

the desired ﬁeld of view (equivalent to 10 HPF) is be-

ing applied. Next, a thresholding with 0.95 is applied

to retain only areas that are covered to at least 95 %

with tissue, resulting in the valid mask V .

Lastly, both maps M and V are used to ﬁnd the

position of the maximum value, constrained to im-

age areas where the valid mask is nonzero. We ex-

pect that these coordinates represent the center of ten

high power ﬁelds with the highest mitotic count of the

WSI.

3.2 Comparison of Two Network

Architectures

Ronneberger’s U-Net architecture (Ronneberger

et al., 2015) has been successfully used in a large

number of segmentation tasks throughout med-

ical imaging, such as aortic stent segmentation

(Breininger et al., 2018), organ segmentation (Chen

et al., 2018) or bone and tumor segmentation (Kay-

alibay et al., 2017). We have shown previously

(Aubreville et al., 2018b), that this architecture can

also be used for direct mitotic ﬁgure segmentation.

However, in this approach, we generate a ﬁne (i.e.

in the same resolution as the original image) segmen-

tation map of the image, where a much more coarse

version of the same map would be sufﬁcient for the

BIOIMAGING 2019 - 6th International Conference on Bioimaging

WSI patch

Mitosis Det. CNN

prediction

ground truth

IoU

conv 3x3x8

conv 3x3x16

conv

3x3x32

conv

3x3x32

conv

3x3x64

conv

3x3x64

conv 3x3x128 conv 3x3x128

512x512x3

32x32x1

conv 1x1x1

conv 1x1x4

32x32x1

pool

Figure 2: Overview of the coarse mitosis detection network (CMDN) and its training. The network predicts a 32×32 map,

i.e. a subsampling of 16, where mitotic ﬁgures are represented by ﬁlled circles. Intersection-over-Union (IoU) is used for

optimization.

WSI patch

Mitosis Det. CNN

prediction

ground truth

IoU

2x conv

3x3x32

2x conv

3x3x64

2x c 3x3x128

512x512x3

512x512x1

2x conv

3x3x64

2x conv

3x3x32

2x conv 3x3x8

pool

2x conv 3x3x16

pool

2x conv 3x3x16

2x conv 3x3x8

conv 1x1x1

conv 1x1x4

Figure 3: Overview of using U-Net as a mitosis detector. The network predicts a 512×512 full-resolution map, where mitotic

ﬁgures are represented by ﬁlled circles. Intersection-over-Union (IoU) is used for optimization.

subsequent steps of forming a map of mitotic density

estimates. We thus investigated the question if the

downsampling path of the U-Net architecture might

be sufﬁcient for the given task, in effect removing the

complete upsampling path with its skip connections

and adding a simple 1×1 convolution layer. We will

denote this approach in the following as coarse mito-

sis detection network (CMDN).

3.2.1 Coarse Mitosis Detection Network

The coarse network (see Fig. 2) consists of 5 stages

of pairs of 2D convolution layers (ﬁlter kernel size:

3x3) followed by a maximum pooling operation (ﬁl-

ter kernel size: 2x2) each. As in the approach by

Ronneberger et al., the ﬁlter depth (or number of ﬁl-

ter channels) doubles with each layer. A 1x1 convo-

lution is being used at the input of the network for

colour space adjustment, and another 1x1 convolu-

tion to generate the output mask with a dimension of

32 × 32 × 1. Batch normalization and rectifying lin-

ear units (ReLU) as nonlinearities are used after each

convolutional layer. The ﬁnal convolution layer uses

a sigmoid activation function. As described in earlier

works (Aubreville et al., 2018b), we utilize negative

Intersection over Union (IoU) as a loss function for

the task, and we minimize this using Adam Optimizer

(Kingma and Ba, 2014) with Tensorﬂow. The use of

IoU as a loss function, as proposed by Rahman and

Wang (Rahman and Wang, 2016), has the advantage

of helping with the strong imbalance problem intro-

duced by the sparsity of mitotic ﬁgures in WSI. The

IOU operator is being applied on the network output

and a ground truth estimate of mitotic ﬁgures (see Fig.

2). Here, since in the subsampled map, centers of mi-

totic ﬁgures will typically not be in the center of the

sampling grid, we use sub-coordinate drawing of the

mask (using the shift parameter of OpenCV’s circle

operation). This results in a more acurrate downsam-

pled representation of the mitotic ﬁgure mask.

3.2.2 U-Net as Mitosis Detection Network

For comparison, we segment the same input images

with a full-resolution mitotic ﬁgure map using Ron-

neberger’s U-Net approach (Ronneberger et al., 2015)

(see Fig. 3). We assume that, as in other evaluations,

the skip-connections between the downsampling path

and the respective same resolution of the upsamling

path will help the network to ﬁnd more accurate re-

sults. Admittedly, this network will have approxi-

mately twice the parameters of the original network,

so it could potentially perform better due to its bigger

capacity.

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks

3.2.3 Training of the Networks

Both networks have been trained for the exact same

number of iterations. We observed that for both net-

works, the training had converged, as visible in a sta-

ble validation loss. For both networks, training sam-

ples were drawn randomly from the complete train-

ing set consisting of 22 Whole Slide Images. In these

training images, the upper 80% was used for training,

while the lower 20% was used for validation.

We employed a strategy, where in each mini-batch

of three images, one image would contain at least one

mitotic ﬁgure, another would be drawn completely

randomly, and one would be the hard example pick,

containing at least one cell where either the experts

did not agree on being a mitotic ﬁgure or it being clas-

siﬁed as mitotic-ﬁgure-similar but not being a mitotic

ﬁgure. Each of these images was taken as a crop with

random rotation from the original WSI. For valida-

tion, images were drawn completely at random from

the respective image region in order to be statistically

as close as possible to the actual test set. Due to this

approach of random sampling, we were not able to

determine a training epoch as by the classical deﬁni-

tion of the network having seen all training images

once. Thus, we consider a run of 15,000 image iter-

ations a pseudo-epoch. As our validation set is com-

paratively large, we chose a random pick of 6,000 im-

ages to be run after each epoch to evaluate the per-

formance. Both networks have been trained for 150

pseudo-epochs.

4 RESULTS

Evaluating both approaches, we ﬁnd a much higher

correlation coefﬁcient between the ground truth mi-

totic count map and the estimated map when using

the U-Net architecture (r = 0.936) compared to the

coarse CMDN approach (r = 0.829). As visible from

Fig. 5 the CMDN had a tendency to overestimate mi-

totic activity in the slides.

This also reﬂects in an overall better performance

in predicting a proper ﬁeld of interest, as seen in

Fig. 4. Although for most test slides, the differences

were minor, we ﬁnd a better region proposal for some

slides, e.g. for test slide 3, which is a relevant bor-

derline slide. As visible in Fig. 6(a), this slide has a

rather unequal distribution of mitotic ﬁgures (and thus

of MC) in the tissue. For all approaches, however, the

position chosen in the relevant slides (3-9), yields a

value in the upper quartile of mitotic count distribu-

tion (Fig. 4).

The evaluation of individual slides (Fig. 7 and Ta-

Figure 4: Box-whisker plots of mitotic count in all slides

of our test data set. Only tumor tissue was included in

this analysis. The arrows indicate the mitotic count (as of

ground truth labels) of the proposed position by the U-Net

detector (red) and the CMDN detector (blue) in the slide,

i.e. the closer it is to the maximum value of the distribution,

the better the estimate.

Figure 5: Relationship between ground truth mitotic count

and estimate for approaches using the coarse network

(CMDN, blue) and the U-Net (red) as mitotic ﬁgure detec-

tion network. Clearly, using the U-Net architecture leads

to an overall better correlation to the ground truth mitotic

count.

ble 1) shows, that correlation between mitotic count

estimate and ground truth is rather weak for slides

with very low mitotic activity (test slides 0 to 2).

Here, both networks tend to overestimate the presence

of mitotic ﬁgures. For borderline (3 to 5) and slides

with high mitotic activity (6 to 9), the correlation is

generally good.

For all individual test slides, the proposed region

reﬂects a region of high mitotic activity on the given

WSI.

BIOIMAGING 2019 - 6th International Conference on Bioimaging

(a) Test slide 3 (b) Test slide 4 (c) Test slide 5

Figure 6: Ground truth distribution of MC (green overlay, where higher opacity indicates higher MC) and region proposal of

the approaches using U-Net (red) and CMDN (blue) for three slides of our test set. For all three slides, the proper choice of

ﬁeld of view has a high inﬂuence on the prognostication.

Table 1: Correlation coefﬁcients between estimated and ground truth mitotic count (MC) in the complete WSI for individual

slides in the test set. While for many slides, performance between both approaches is comparable, slide 0, 4, 6 and 9 yield a

clear advantage for the approach utilizing U-Net.

Test slide 0 1 2 3 4 5 6 7 8 9

CMDN 0.545 0.019 0.354 0.895 0.713 0.932 0.858 0.971 0.974 0.911

U-Net 0.617 0.111 0.322 0.848 0.872 0.919 0.924 0.976 0.983 0.973

Figure 7: Semitransparent scatter plots of individual test slides with the U-Net detector in red and the CMDN detector in

blue. As obvious from slides 3,4 and 5, the CMDN detector tends to overestimate the mitotic count especially in slides with

borderline specimen.

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks

5 DISCUSSION

We demonstrated that, while the general problem of

identifying mitotic ﬁgures in whole slide images with

high accuracy, is still far from being achieved, the

outcomes of mitosis detection approaches might well

serve as an intermediate step. In preselecting the ﬁeld

of interest containing the highest mitotic ﬁgure den-

sity, the algorithm can help the pathologist in deter-

mining the area of the tumor where the proliferation

is the highest. Hence, we expect that such approaches

can lead to a more reproducible grading and thus po-

tentially better tailored treatment of the patient.

The most crucial slides for the approach are slides

3 to 5, as also shown in Fig. 6(a) to 6(c). Because

the mitotic ﬁgure distribution in these slides is rather

patchy i.e. with strong regional differences (see also

Fig. 4 for absolute numbers) an arbitrary selection is

likely to not yield the area with highest mitotic count,

and thus the grading is subject to a possible strong

variance. For all of these cases, both approaches were

well able to pick an area with very high mitotic ac-

tivity, with only minor differences in performance.

The U-Net approach, while leading to a considerably

higher correlation coefﬁcient on the overall data set,

did not lead to a signiﬁcantly better overall perfor-

mance.

Our approach did not employ stain normalization

methods, as done in the majority of recent mitosis de-

tection approaches (Veta et al., 2018). This was done

in part, because the staining quality of our dataset is

relatively stable due to the usage of a tissue stainer

(ST5010 Autostainer XL, Leica, Germany) and all

slides being created and scanned in the same lab. Ad-

ditionally, we assume that with the high number of

included WSI in the present study, natural variance of

stain becomes less relavant. For application of this

approach on another (possibly smaller) data set, how-

ever, we would recommend investigating a positive

inﬂuence of such methods.

It is important to state that the results of this work

were achieved on a limited test data set for canine

mast cell tumors. While, theoretically, we would not

assume different performance on different tumors, tis-

sues or species, this should certainly be investigated.

Another important question is to what degree the im-

proved stability of region proposal, as shown in this

work, would lead to a lower inter-rater-variability in

grading, which we aim to deal with in future work.

REFERENCES

Aubreville, M., Bertram, C., Klopﬂeisch, R., and Maier, A.

(2018a). SlideRunner - A Tool for Massive Cell Anno-

tations in Whole Slide Images. In Maier, A., Deserno,

T. M., Handels, H., Maier-Hein, K. H., Palm, C., and

Tolxdorff, T., editors, Bildverarbeitung f

ur die Medi-

zin 2018. Proceedings des Workshops vom 11. bis 13.

arz 2018 in Erlangen, pages 309–314.

Aubreville, M., Bertram, C. A., Klopﬂeisch, R., and Maier,

A. (2018b). Augmented mitotic cell count using ﬁeld

of interest proposal. Bildverarbeitung f

ur die Medizin

2019 (accepted), arXiv:1810.00850.

Baak, J. P. A., Gudlaugsson, E., Skaland, I., Guo, L.

H. R., Klos, J., Lende, T. H., Søiland, H., Janssen, E.

A. M., and zur Hausen, A. (2008). Proliferation is the

strongest prognosticator in node-negative breast can-

cer: signiﬁcance, error sources, alternatives and com-

parison with molecular prognostic markers. Breast

Cancer Research and Treatment, 115(2):241–254.

Boiesen, P., Bendahl, P. O., Anagnostaki, L., Domanski,

H., Holm, E., Idvall, I., Johansson, S., Ljungberg, O.,

Ringberg, A.,

Ostberg, G., and Fern

o, M. (2000). His-

tologic grading in breast cancer: reproducibility be-

tween seven pathologic departments. Acta Oncolog-

ica, 39(1):41–45.

Bonert, M. and Tate, A. J. (2017). Mitotic counts in breast

cancer should be standardized with a uniform sample

area. BioMedical Engineering OnLine, 16(1):28.

Breininger, K., Albarqouni, S., Kurzendorfer, T., Pﬁster,

M., Kowarschik, M., and Maier, A. K. (2018). Intra-

operative stent segmentation in X-ray ﬂuoroscopy for

endovascular aortic repair. Int. J. Computer Assisted

Radiology and Surgery.

Chen, S., Roth, H., Dorn, S., May, M., Cavallaro, A., Lell,

M., Kachelrie

s, M., Oda, H., Mori, K., and Maier, A.

(2018). Towards Automatic Abdominal Multi-Organ

Segmentation in Dual Energy CT using Cascaded 3D

Fully Convolutional Network. In Noo, F., editor, the

ﬁfth edition of The International Conference on Im-

age Formation in X-ray Computed Tomography, pages

395–398.

Cires¸an, D. C., Giusti, A., Gambardella, L. M., and Schmid-

huber, J. (2013). Mitosis detection in breast can-

cer histology images with deep neural networks. In-

ternational Conference on Medical Image Computing

and Computer-Assisted Intervention (MICCAI), 16(Pt

2):411–418.

Edmondson, E. F., Hess, A. M., and Powers, B. E. (2014).

Prognostic Signiﬁcance of Histologic Features in Ca-

nine Renal Cell Carcinomas. Veterinary Pathology,

52(2):260–268.

Elston, C. W. and Ellis, I. O. (1991). Pathological prog-

nostic factors in breast cancer. i. the value of his-

tological grade in breast cancer: experience from a

large study with long-term follow-up. Histopathology,

19(5):403–410.

Horobin, R. W. (1969). The impurities of biological dyes:

their detection, removal, occurrence and histological

signiﬁcance? A review. The Histochemical Journal,

1(3):231–265.

BIOIMAGING 2019 - 6th International Conference on Bioimaging

Kayalibay, B., Jensen, G., and van der Smagt, P. (2017).

Cnn-based segmentation of medical imaging data.

arXiv preprint arXiv:1701.03056.

Kingma, D. P. and Ba, J. (2014). Adam: A method for

stochastic optimization.

Li, C., Wang, X., Liu, W., and Latecki, L. J. (2018). Deep-

Mitosis: Mitosis detection via deep detection, veri-

ﬁcation and segmentation networks. Medical Image

Analysis, 45:121–133.

Martin, A. R., Weisenburger, D. D., Chan, W. C., Ruby,

E. I., Anderson, J. R., Vose, J. M., Bierman, P. J., Bast,

M. A., Daley, D. T., and Armitage, J. O. (1995). Prog-

nostic value of cellular proliferation and histologic

grade in follicular lymphoma. Blood, 85(12):3671–

3678.

Meuten, D. J., Moore, F. M., and George, J. W. (2016). Mi-

totic Count and the Field of View Area. Veterinary

Pathology, 53(1):7–9.

Otsu, N. (1979). A Threshold Selection Method from Gray-

Level Histograms. IEEE Transactions on Systems,

Man, and Cybernetics, 9(1):62–66.

Rahman, M. A. and Wang, Y. (2016). Optimizing

intersection-over-union in deep neural networks for

image segmentation. In International Symposium on

Visual Computing, pages 234–244. Springer.

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

- Convolutional Networks for Biomedical Image Seg-

mentation. International Conference on Medical Im-

age Computing and Computer-Assisted Intervention

(MICCAI), 9351(Chapter 28):234–241.

Roux, L., Racoceanu, D., Lom

enie, N., Kulikova, M.,

Irshad, H., Klossa, J., Capron, F., Genestie, C.,

Le Naour, G., and Gurcan, M. N. (2013). Mitosis de-

tection in breast cancer histological images An ICPR

2012 contest. Journal of pathology informatics, 4:8.

Van Diest, P. J., Baak, J. P. A., Matze-Cok, P., Wisse-

Brekelmans, E. C. M., van Galen, C. M., Kurver, P.

H. J., Bellot, S. M., Fijnheer, J., van Gorp, L. H. M.,

Kwee, W. S., Los, J., Peterse, J. L., Ruitenberg, H. M.,

Schapers, R. F. M., Schipper, M. E. I., Somsen, J. G.,

Willig, A. W. P. M., and Ariens, A. T. (1992). Repro-

ducibility of mitosis counting in 2,469 breast cancer

specimens: Results from the Multicenter Morphome-

tric Mammary Carcinoma Project. Human Pathology,

23(6):603–607.

Veta, M., Heng, Y. J., Stathonikos, N., Bejnordi, B. E.,

Beca, F., Wollmann, T., Rohr, K., Shah, M. A., Wang,

D., Rousson, M., et al. (2018). Predicting breast tumor

proliferation from whole-slide images: the tupac16

challenge. arXiv preprint arXiv:1807.08284.

Veta, M., van Diest, P. J., Willems, S. M., Wang, H., Mad-

abhushi, A., Cruz-Roa, A., Gonzalez, F., Larsen, A.

B. L., Vestergaard, J. S., Dahl, A. B., Cires¸an, D. C.,

Schmidhuber, J., Giusti, A., Gambardella, L. M.,

Tek, F. B., Walter, T., Wang, C.-W., Kondo, S., Ma-

tuszewski, B. J., Precioso, F., Snell, V., Kittler, J.,

de Campos, T. E., Khan, A. M., Rajpoot, N. M.,

Arkoumani, E., Lacle, M. M., Viergever, M. A., and

Pluim, J. P. W. (2015). Assessment of algorithms for

mitosis detection in breast cancer histopathology im-

ages. Medical Image Analysis, 20(1):237–248.

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks