Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy

Images - A Cross-site Robustness Assessment

Marc Aubreville

, Miguel Goncalves

, Christian Knipfer

3,4

, Nicolai Oetter

5,4

, Tobias W

urﬂ

Helmut Neumann

, Florian Stelzle

5,4

, Christopher Bohr

and Andreas Maier

1,4

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

at Erlangen-N

urnberg, Germany

Department of Otorhinolaryngology, Head and Neck Surgery, University Hospital Erlangen,

Friedrich-Alexander-Universit

at Erlangen-N

urnberg, Germany

Department of Oral and Maxillofacial Surgery, University Medical Center Hamburg-Eppendorf, Germany

Erlangen Graduate School in Advanced Optical Technologies (SAOT),

Friedrich-Alexander-Universit

at Erlangen-N

urnberg, Germany

Department of Oral and Maxillofacial Surgery, University Hospital Erlangen,

Friedrich-Alexander-Universit

at Erlangen-N

urnberg, Germany

First Department of Internal Medicine, University Medical Center Mainz,

Johannes Gutenberg-Universit

at Mainz, Germany

Keywords:

Automatic Carcinoma Detection, Confocal Laser Endomicroscopy, Deep Convolutional Networks, Squamous

Cell Carcinoma.

Abstract:

Deep learning technologies such as convolutional neural networks (CNN) provide powerful methods for image

recognition and have recently been employed in the ﬁeld of automated carcinoma detection in confocal laser

endomicroscopy (CLE) images. CLE is a (sub-)surface microscopic imaging technique that reaches magniﬁ-

cations of up to 1000x and is thus suitable for in vivo structural tissue analysis.

In this work, we aim to evaluate the prospects of a priorly developed deep learning-based algorithm targeted

at the identiﬁcation of oral squamous cell carcinoma with regard to its generalization to further anatomic loca-

tions of squamous cell carcinomas in the area of head and neck. We applied the algorithm on images acquired

from the vocal fold area of ﬁve patients with histologically veriﬁed squamous cell carcinoma and presumably

healthy control images of the clinically normal contra-lateral vocal cord.

We ﬁnd that the network trained on the oral cavity data reaches an accuracy of 89.45% and an area-under-

the-curve (AUC) value of 0.955, when applied on the vocal cords data. Compared to the state of the art, we

achieve very similar results, yet with an algorithm that was trained on a completely disjunct data set. Concate-

nating both data sets yielded further improvements in cross-validation with an accuracy of 90.81% and AUC

of 0.970.

In this study, for the ﬁrst time to our knowledge, a deep learning mechanism for the identiﬁcation of oral

carcinomas using CLE Images could be applied to other disciplines in the area of head and neck. This study

shows the prospect of the algorithmic approach to generalize well on other malignant entities of the head and

neck, regardless of the anatomical location and furthermore in an examiner-independent manner.

1 INTRODUCTION

Squamous cell carcinoma is a common kind of cancer,

found in epithelial tissue. The prevalence within the

head and neck region is estimated to be around 1.3

million cases per year (Forastiere et al., 2001; Ferlay

et al., 2014).

Many cases of head and neck squamous cell carci-

noma (HNSCC) are diagnosed at a late stage, which

impairs treatment outcomes and increases mortality

(Muto et al., 2004). The gold standard of diagno-

sis is invasive biopsy of the tissue with subsequent

histopathological assessment (Oetter et al., 2016).

However, biopsies carry the risk of infections and

bleeding. Furthermore, due to the invasiveness a limi-

tation in the sample size and quantity hinders the ﬁnd-

ing of accurate resection margins (Dittberner et al.,

2016; Nathan et al., 2014). An non- or minimally

invasive in vivo characterization of microstructures

could help detecting such malignancies at an early

stage while at the same time reducing risk. Further,

it could be of help for periodic monitoring of possibly

malignant cellular structures, reducing the risk for un-

necessary biopsies.

Aubreville, M., Goncalves, M., Knipfer, C., Oetter, N., Würﬂ, T., Neumann, H., Stelzle, F., Bohr, C. and Maier, A.

Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment.

DOI: 10.5220/0006534700270034

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 2: BIOIMAGING, pages 27-34

ISBN: 978-989-758-278-3

Hard palate

(OC)

Inner Labium (OC)

Alveolar ridge

(OC)

Vocal co

(VC)

Figure 1: Anatomical locations from the oral cavity and the

upper aero-digestive and respiratory tract.

One method that has successfully been applied

for visual inspection of suspicious lesions is Con-

focal Laser Endomicroscopy (CLE). In this imaging

method, laser light is emitted and applied on tis-

sue using a ﬁbre-optic bundle that is typically in-

serted through the accessory channel of an endoscope

(Chauhan et al., 2014). The resolution of CLE is high,

providing magniﬁcations of up to 1000x (Oetter et al.,

2016) and enabling sub-cellular imaging. A contrast

agent (ﬂuorescein) is applied intravenously prior to

the examination in order to stain the intercellular gap

and hence outline the cell borders. CLE is success-

fully used in clinical routine diagnostics of the intes-

tine (Neumann et al., 2010) and was recently also suc-

cessfully applied on cancer assessment in the oral cav-

ity (Oetter et al., 2016) and the upper respiratory tract

(Goncalves et al., 2017).

Due to the property of making cellular structures

visible, CLE is said to provide ’real-time’ optical

biopsies (Parikh et al., 2016), which is a major ad-

vantage over the need to perform traditional biopsies,

e.g. when ﬁnding a proper resection margin for intra-

operative monitoring during surgical tumor removal.

However, it was shown that the accuracy in interpreta-

tion of CLE images is highly dependent on the experi-

ence of the clinical expert, and that a signiﬁcant learn-

ing curve exists (Neumann et al., 2011). An automatic

detection and interpretation of such images could thus

help to improve the standard and make CLE also ap-

plicable with less training involved.

Deep learning methods, such as convolutional

neural networks (CNN) have recently been used in a

variety of image recognition tasks. We have shown

that CNN-based recognition methods outperform the

state of the art in HNSCC detection on CLE im-

ages, using a data set of 12 patients (Aubreville et al.,

2017). In order to investigate the robustness of the

method, generalization has to be assessed to other en-

vironments. One step into showing this generalization

is to apply a trained machine learning model from one

anatomical site and clinical team to another, without

any modiﬁcation of the underlying model structure

and content. This would provide a strong hint of gen-

eralization towards other locations of the upper aero-

digestive tract with similar but not identical epithelia.

2 MATERIAL

For the present work, we are using images from two

anatomical locations (see ﬁgure 1): From within the

oral cavity, we used images from three clinically nor-

mal sites and a lesion site with veriﬁed SCC. From

the upper aero-digestive and respiratory tract, we used

images of the vocal cords (clinically normal and with

veriﬁed malign changes). All images were acquired

using a probe-based CLE (pCLE) system (Cellvizio,

Mauna Kea Technologies, Paris, France). From all

patients, written informed consent was obtained prior

to the study. Approval was granted by the respec-

tive institutional review boards. The research was car-

ried out in accordance with the Code of Ethics of the

World Medical Association (Declaration of Helsinki)

and the guidelines of the Friedrich-Alexander Univer-

sity Erlangen-Nuremberg.

2.1 Oral Cavity (OC)

We included image sequences (N = 116) from 12 pa-

tients with diagnosed and veriﬁed HNSCC in the oral

cavity that were recorded at the Department of Oral

and Maxillofacial Surgery (University Hospital Er-

langen)

. For all patients, imaging was performed at

the suspected carcinoma site, as well as three other

anatomical sites within the oral cavity. After veriﬁ-

cation of the carcinoma diagnosis by histo-pathology,

all patients underwent surgery for removal of the sus-

pected tissue. Oetter et al. found that the accuracy

as rated by CLE-experienced specialists was 92.3%,

where the experts were allowed to see the complete

video sequence to base their assessment upon. So,

this number accounts for occurrence of singular or

sparse cues for a correct classiﬁcation. It is to be ex-

pected that an assessment of a whole sequence can

achieve a better performance than on singular images.

In contrast, we performed evaluation on a per-frame

base. The total number of images with sufﬁcient qual-

ity in this data set is 7,894.

Study approved by the ethics committee of the Univer-

sity of Erlangen-N

urnberg; reference number: 243 12 B

BIOIMAGING 2018 - 5th International Conference on Bioimaging

Figure 2: Overview of the classiﬁcation method. Patches (1,2,...) are extracted from the visible view area of the image and

independently fed into a CNN-based patch classiﬁcation. The resulting probabilities are then fused in a ﬁnal step to one

class-probability tuple for the whole image.

2.2 Vocal Cords (VC)

Squamous cell carcinoma of the vocal cords is the

most prevalent form of cancer of the laryngeal tract

(Parkin et al., 2005) and an anatomy where bright

light endoscopy alone is insufﬁcient for diagnosis,

because a signiﬁcant range of benign mucosal al-

terations with similar macroscopic appearance exist

(Goncalves et al., 2017). Biopsies require a sufﬁ-

cient amount of material and extensive biopsy in this

anatomical region causes functional problems, such

as chronic hoarseness or other voice modiﬁcations

(Cikojevi

c et al., 2008). Accurate diagnosis using op-

tical, non-invasive methods would thus provide a sig-

niﬁcant improvement for the life quality of patients

with suspected cancer of the upper respiratory tract.

In this study, we included N = 73 image sequences

from ﬁve patients with veriﬁed malignant transfor-

mations of the vocal folds. For each patient, the

contra-lateral vocal cord (that was clinically normal

in all cases) was also investigated using CLE, build-

ing a presumably healthy data set. The images were

acquired at the Department of Otorhinolaryngology,

Head and Neck Surgery, University Hospital Erlan-

gen

. All image sequences were taken during micro-

laryngoscopy and were pre-selected by a clinical ex-

pert in CLE imaging, in order to remove images

where the acquisition was not performed properly,

e.g. when the probe was not in contact with the tissue.

For one patient only image sequences of the clinically

normal vocal cord exist in a sufﬁcient quality, reduc-

ing the number of carcinoma patients to four. The

total number of images in this data set is 4,425. Pre-

vious studies of automatic carcinoma detection on this

data set have shown accuracies in grading of between

86.4% and 89.8% (Vo et al., 2017), using the methods

of Jaremenko (Jaremenko et al., 2015).

Study approved by the ethics committee of the Univer-

sity of Erlangen-N

urnberg; reference number: 60 14 B

Goncalves et al. selected 31 representative images

out of a different vocal cord data set of 7 patients and

found rating accuracies by ENT specialists to be be-

tween 58.1% and 87.1%, where the non-CLE expe-

rienced doctors had a mean accuracy of 67.7% and

those with profound experience of CLE one of 82.2%

(Goncalves et al., 2017). Also for this data set, the

differences in recognition performance are well ex-

plained by the learning effect as described by Neu-

mann (Neumann et al., 2011). However, the rela-

tively low accuracies for less experienced observers

also motivate the development of an automated ap-

proach, as pointed out by Goncalves.

2.3 Image Quality and Artifact

Occurrence

CLE images suffer from a range of deteriorations

in image quality (Neumann et al., 2012; Aubreville

et al., 2017). While some images are tainted by mo-

tion artifacts, others have a very low signal to noise

ratio (see ﬁgure 4, second column).

For images acquired by CLE, image noise inten-

sity is mostly dependent on the actual amount of re-

ceived photons, i.e. the optical response of the tissue,

since the optical receiver has a broad dynamic range

and images are automatically range compressed by

the viewing software to ﬁt the gray level range (8bit)

of today’s screens and image ﬁle formats. One met-

ric to measure this optical response is by the pixel

value distribution of the image. We calculated the me-

dian pixel value of each image to assess the optical

response and compared the different anatomical loca-

tions of clinically normal tissue (see ﬁgure 5). CLE

images have sparse occurrences of vessels (see 4, top

left), which are typically very bright, which is why the

median might be beneﬁcial for contrast assessment of

the cellular structures of the overall image. Low val-

ues of the median indicate a noisy image, while high

Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment

clinically normal vocal fold

veriﬁed carcinoma

Figure 3: Selected CLE Images acquired from the vocal fold area. On the left, presumably healthy images are depicted that

were taken from the clinically normal contralateral vocal cord of patients with epithelial cancer one of the both vocal cord.

clinically normal oral cavity images

veriﬁed carcinoma

labial

hard palate

alveolar ridge

Figure 4: Selected CLE Images acquired from the oral cavity. On the left, presumably healthy images are depicted that

were taken from clinically normal regions from patients suffered from histologically veriﬁed head and neck squamous cell

carcinoma.

Table 1: Test conditions. For cross-correlation, a leave-one-patient out scheme was used.

Name Train data set Patients training Test data set Patients test Cross validation

OC Oral cavity 11 Oral cavity 1 LOPO

VC Vocal cords 4 Vocal cords 1 LOPO

OC/VC Oral cavity 12 Vocal cords 5 -

VC/OC Vocal cords 5 Oral cavity 12 -

OC+VC both 16 both 1 LOPO

values indicate an image with a high optical response

to the laser. For low values, the range compression

of the viewing system is in fact an extension, which

ampliﬁes noise further for the examiner.

We ﬁnd that a signiﬁcant amount of images of the

palatinal and labial region have a low response, indi-

cating high noise levels, which is also conﬁrmed by

optical inspection of the images. In contrast, images

BIOIMAGING 2018 - 5th International Conference on Bioimaging

Figure 5: Normalized histogram of the median value for the

different classes of clinically normal tissue from both data

sets. Due to the wide range of pixel values, the histogram is

given at log scale.

of the alveolar ridge and - even more - the vocal folds

tend to have a better signal to noise ratio, as depicted

in ﬁgure 5.

This can be related to the different anatomical

properties of the respective epithelia: Regions with

high mechanical stress due to chewing have a higher

degree of corniﬁcation (Rohen and L

utjen-Drecoll,

2000). Speciﬁcally, the hard palate is known to have

a high degree of corniﬁcation (Luellmann-Rauch,

2015). The inner lip (labium) is generally not consid-

ered a corniﬁed epithelium, however our images were

taken at the intersection between mucous membrane

and outer lip with its epidermal layer, where corniﬁ-

cation is indeed prominent (Rohen, 1994). Contrary

to this, the vocal cords are known to consist of mul-

tiple layers of uncorniﬁed squamous epithelium (Ro-

hen, 1994).

The difference in image quality could, however,

also be caused by a different preselection bias be-

tween the two clinical teams.

In our data sets, motion artifact incidence is signif-

icantly higher in the oral cavity images compared to

the vocal cords images. This can be related to patients

being under general anaesthesia in case of the vocal

fold microendoscopy (Goncalves et al., 2017). In this

case, the only reason for motion is hand movement of

the clinician performing the image acquisition.

3 METHODS

Our method is based on the extraction of squared

patches from the round ﬁeld of view area of a CLE

image, classiﬁcation by a deep convolutional network

(CNN). Subsequently, the a posteriori probabilities

are fused (see Fig. 2; for more details, see (Aubre-

ville et al., 2017)). The approach limits overﬁtting

of the CNN model by a small patch size (80x80 px)

and thus a reduced capacity of the network. Addition-

ally, the strategy results in a large number of training

samples, since every image consists of a multitude of

patches. We trained the network for 60 epochs, us-

ing the Adam optimizer at an initial step size of 0.01

within the TensorFlow framework.

In total, we performed two additional test sets:

1. Generalization Tests

We performed training of our deep convolutional

models on one anatomical location and testing on

the other (tests OC/VC and VC/OC, see table 1).

2. Algorithmic Validation

We performed a validation of the automatic de-

tection algorithm (Aubreville et al., 2017) on the

vocal cords data set and on the concatenated data

set (tests VC and OC+VC, respectively).

For all tests, where train and test data were

taken from the same data set, we applied leave-

one-patient-out cross validation. Independent x-fold

cross-validation or simple random train-test-splitting

isn’t applicable, since high correlations between con-

secutive frames within one sequence might exist. The

VC data set is small compared to the other data sets,

has a much lower number of patients and comes from

a small anatomical structure. This leads us to expect

generalization to the oral cavity to work better than

vice versa.

Our intention for the last test (OC+VC) is, how

well the algorithm is able to improve overall results

from more image material.

4 RESULTS

We ﬁnd that the patch-based classiﬁcation method

seems to generalize well from the oral cavity data set

to the vocal cords dataset (ROC area-under-the-curve

of 0.9548). It is a comparable ﬁgure to the original

data, where the ROC AUC was 0.9550. Trained on

the vocal cords data set, the method outperforms the

cross-validation results as reported by Vo et al. for the

method by Jaremenko et al. slightly (Vo et al., 2017;

Jaremenko et al., 2015). When comparing the results

on individual patients, slight differences between both

approaches occur, where the approach only trained on

VC data performs better for patient 3, while the ap-

proach trained on OC data performs better for patient

Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment

Table 2: Results of all tests. For the cross validation cases OC, VC and OC+VC, the results were calculated on the concate-

nated result vector of all cross validation steps.

Condition Accuracy Precision Recall ROC area under curve

OC (Aubreville et al., 2017) 88.34% 85.40% 91.10% 0.9550

VC 91.39% 93.64% 92.03% 0.9484

OC/VC 89.45% 87.47% 96.37% 0.9548

VC/OC 68.53% 60.81% 95.63% 0.8484

OC+VC 90.81% 90.12% 92.59% 0.9697

Figure 6: Receiver Operating Characteristic (ROC) curve

for the different setups. (OC=oral cavity, VC=vocal cords).

1. The concatenated data set increased performance

for all patients in cross-validation (cf. ﬁgure 7).

The generalization task from the vocal cord data

set to the much larger oral cavity data set, however,

did not show comparable results, having AUC val-

ues of only 0.8484. Inspecting individual patient per-

formance, it is obvious that the generalization loss is

prominent in a number of patients, while others, like

the tests on patient 4,5,6 and 10 perform comparable

to the tests on the original OC data set.

When the data set is concatenated (condition

OC+VC), the accuracy and ROC AUC values in-

creases, with values of 90.81% and 0.9697, respec-

tively.

5 DISCUSSION

The much greater variance of the oral cavity data set

(cf. ﬁgure 5) due to the larger variety in acquisition

conditions led to better generalization properties com-

pared to the vocal cords data set. The general signal-

to-noise ratio was much better in this case. This is the

reason why the classiﬁer trained on the vocal cords

tends to confuse noisier images, as they have been

recorded from the corniﬁed sections of the oral cavity,

for malignant tissue. This is also reﬂected by the high

recall and low precision ratings in this classiﬁcation

Figure 7: Accuracy for all patients with both classes of the

vocal fold data set.

Figure 8: Accuracy for all patients of the oral cavity data

set.

task (cf. 4th row of table 2).

For the generalization from the oral cavity to the

vocal fold data set, this restriction did not apply, since

the CLE imaging conditions within the oral cavity

seems to be a superset of those on the vocal folds.

However, also the greater number of patients for

whom veriﬁed carcinoma imaging material was avail-

able likely played a role, which is also indicated by

the increased performance in cross-validation for the

concatenated data set. This indicates that the pattern

recognition capacity of the model is not yet reached

and that additional imaging data would likely increase

BIOIMAGING 2018 - 5th International Conference on Bioimaging

performance further.

Since histological veriﬁcation was only present

for cancerous areas in both data sets, we can only as-

sume that clinically normal regions represent healthy

epithelium. Extraction of tissue from those regions

would however be ethically questionable and not re-

ceive approval of the review boards.

It is questionable, if an 100% accurate classiﬁca-

tion of epithelial tissue is possible using CLE alone,

as even experts in the ﬁeld of CLE were not able to

classify cancerous tissue perfectly (Goncalves et al.,

2017; Oetter et al., 2016). Due to the low penetration

depth of CLE, it is sometimes possible to overlook tu-

mors that spread within the submucosa. Such tumors

could be visualized only through histological section

or perhaps through Optical Coherence Tomography

(Betz et al., 2015).

One important aspect in automated inspection of

CLE images is the removal of artifact-tainted images

prior to training, since artifact occurrence is corre-

lated with the surface conditions of the epithelium.

This implies that it is also correlated to the malig-

nancy classiﬁcation, a causal relationship between

artifact prevalence and tissue classiﬁcation should,

however, be neglected. This step was done manu-

ally in this work and this problem is subject of future

work.

Even though our approach found good accuracy

ratings, generalization can not be claimed to be fully

shown with this study, due to the limited amount of

patient data. Because of this, future work of our re-

search group will concentrate on the acquisition of

imaging data in order to increase the variance in the

data set, which will presumably increase performance

and robustness of the algorithmic approach.

6 SUMMARY

In this work, we have shown the principal ability to

generalize patch-based CLE image classiﬁcation with

convolutional networks of potentially cancerous ep-

ithelium from a more diversiﬁed data set (from the

oral cavity) to one of another anatomical location (the

vocal folds) with less variance. The second data set

was from a different clinic and a different team.

The generalization showed very promising re-

sults and concatenation of both sets did show fur-

ther improvements in a leave-one-patient-out cross-

validation scenario.

In total, we achieved an accuracy of 89.45% in

the generalization task, where the classiﬁcation model

was trained on the oral cavity data set and applied on

the vocal cords data set. For the concatenated data

set with 17 patients, we achieved a total accuracy of

90.81% for the complete data set.

REFERENCES

Aubreville, M., Knipfer, C., Oetter, N., Jaremenko, C., Rod-

ner, E., Denzler, J., Bohr, C., Neumann, H., Stelzle,

F., and Maier, A. K. (2017). Automatic classiﬁcation

of cancerous tissue in laserendomicroscopy images of

the oral cavity using deep learning. Scientiﬁc Reports

7:11979.

Betz, C. S., Kraft, M., Arens, C., Schuster, M., Pfef-

fer, C., R

uhm, A., Stepp, H., Englhard, A., and

Volgger, V. (2015). Optische Diagnoseverfahren

zur Tumorfr

uhdiagnostik im oberen Luft-Speise-Weg.

HNO, 64(1):41–48.

Chauhan, S. S., Dayyeh, B. K. A., Bhat, Y. M., Gottlieb,

K. T., Hwang, J. H., Komanduri, S., Konda, V., Lo,

S. K., Manfredi, M. A., Maple, J. T., et al. (2014).

Confocal laser endomicroscopy. Gastrointestinal en-

doscopy, 80(6):928–938.

Cikojevi

c, D., Glun

c, I., and Pe

suti

c-Pisac, V. (2008).

Comparison of contact endoscopy and frozen section

histopathology in the intra-operative diagnosis of la-

ryngeal pathology. The Journal of Laryngology &

Otology, 122(8):836–839.

Dittberner, A., Rodner, E., Ortmann, W., Stadler, J.,

Schmidt, C., Petersen, I., Stallmach, A., Denzler, J.,

and Guntinas-Lichius, O. (2016). Automated analy-

sis of confocal laser endomicroscopy images to detect

head and neck cancer. Head & Neck, 38(S1):E1419–

E1426.

Ferlay, J., Soerjomataram, I., Dikshit, R., Eser, S., Mathers,

C., Rebelo, M., Parkin, D. M., Forman, D., and Bray,

F. (2014). Cancer incidence and mortality worldwide:

Sources, methods and major patterns in GLOBOCAN

2012. International Journal of Cancer, 136(5):E359–

E386.

Forastiere, A., Koch, W., Trotti, A., and Sidransky, D.

(2001). Head and Neck Cancer. The New England

Journal of Medicine, 345(26):1890–1900.

Goncalves, M., Iro, H., Dittberner, A., Agaimy, A., and

Bohr, C. (2017). Value of confocal laser endomi-

croscopy in the diagnosis of vocal cord lesions. Eu-

ropean Review for Medical and Pharmacological Sci-

ences, 21(18):3990-3997.

Jaremenko, C., Maier, A., Steidl, S., Hornegger, J., Oet-

ter, N., Knipfer, C., Stelzle, F., and Neumann, H.

(2015). Classiﬁcation of Confocal Laser Endomicro-

scopic Images of the Oral Cavity to Distinguish Patho-

logical from Healthy Tissue. In Bildverarbeitung f

die Medizin 2015, pages 479–485. Springer Berlin

Heidelberg.

Luellmann-Rauch, R. (2015). Taschenbuch Histologie.

Thieme, Stuttgart u.a.

Muto, M., Nakane, M., Katada, C., Sano, Y., Ohtsu, A.,

Esumi, H., Ebihara, S., and Yoshida, S. (2004). Squa-

mous cell carcinoma in situ at oropharyngeal and hy-

Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment

popharyngeal mucosal sites. Cancer, 101(6):1375–

1381.

Nathan, C. A. O., Kaskas, N. M., Ma, X., Chaudhery, S.,

Lian, T., Moore-Medlin, T., Shi, R., and Mehta, V.

(2014). Confocal Laser Endomicroscopy in the De-

tection of Head and Neck Precancerous Lesions. Oto-

laryngology – Head and Neck Surgery, 151(1):73–80.

Neumann, H., Kiesslich, R., Wallace, M. B., and Neurath,

M. F. (2010). Confocal Laser Endomicroscopy: Tech-

nical Advances and Clinical Applications. YGAST,

139(2):388–392.e2.

Neumann, H., Langner, C., Neurath, M. F., and Vieth, M.

(2012). Confocal Laser Endomicroscopy for Diagno-

sis of Barrett’s Esophagus. Frontiers in oncology, 2.

Neumann, H., Vieth, M., Atreya, R., Neurath, M. F., and

Mudter, J. (2011). Prospective evaluation of the learn-

ing curve of confocal laser endomicroscopy in patients

with IBD. Histology and histopathology, 26(7):867–

872.

Oetter, N., Knipfer, C., Rohde, M., Wilmowsky, C., Maier,

A., Brunner, K., Adler, W., Neukam, F.-W., Neumann,

H., and Stelzle, F. (2016). Development and validation

of a classiﬁcation and scoring system for the diagno-

sis of oral squamous cell carcinomas through confo-

cal laser endomicroscopy. Journal of Translational

Medicine, 14(1):1–11.

Parikh, N. D., Gibson, J., Nagar, A., Ahmed, A. A.,

and Aslanian, H. R. (2016). Confocal laser en-

domicroscopy features of sessile serrated adeno-

mas/polyps. United European Gastroenterology Jour-

nal, 4(4):599–603.

Parkin, D. M., Bray, F., Ferlay, J., and Pisani, P. (2005).

Global Cancer Statistics, 2002. CA: A Cancer Journal

for Clinicians, 55(2):74–108.

Rohen, J. W. (1994). Histologische Differentialdiagnose:

Anleitung zur Diagnose histologischer Pr

aparate.

Rohen, J. W. and L

utjen-Drecoll, E. (2000). Funktionelle

Histologie. Schattauer, 4 edition.

Vo, K., Jaremenko, C., Bohr, C., Neumann, H., and Maier,

A. (2017). Automatic Classiﬁcation and Pathological

Staging of Confocal Laser Endomicroscopic Images

of the Vocal Cords. In Maier-Hein, K. H., editor, Bild-

verarbeitung f

ur die Medizin 2017 Algorithmen Sys-

teme Anwendungen, pages 312–317, Heidelberg.

BIOIMAGING 2018 - 5th International Conference on Bioimaging