Approximation of a pipeline of unsupervised retina image analysis methods with a CNN

Friso Heslinga, Josien Pluim, Behdad Dasht Bozorg, Tos Berendschot, Alfons J.H.M. Houben, Ronald M.A. Henry, Mitko Veta

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

9 Downloads (Pure)

Abstract

A pipeline of unsupervised image analysis methods for extraction of geometrical features from retinal fundus images has previously been developed. Features related to vessel caliber, tortuosity and bifurcations, have been identified as potential biomarkers for a variety of diseases, including diabetes and Alzheimer's. The current computationally expensive pipeline takes 24 minutes to process a single image, which impedes implementation in a screening setting. In this work, we approximate the pipeline with a convolutional neural network (CNN) that enables processing of a single image in a few seconds. As an additional benefit, the trained CNN is sensitive to key structures in the retina and can be used as a pretrained network for related disease classification tasks. Our model is based on the ResNet-50 architecture and outputs four biomarkers that describe global properties of the vascular tree in retinal fundus images. Intraclass correlation coefficients between the predictions of the CNN and the results of the pipeline showed strong agreement (0.86 - 0.91) for three of four biomarkers and moderate agreement (0.42) for one biomarker. Class activation maps were created to illustrate the attention of the network. The maps show qualitatively that the activations of the network overlap with the biomarkers of interest, and that the network is able to distinguish venules from arterioles. Moreover, local high and low tortuous regions are clearly identified, confirming that a CNN is sensitive to key structures in the retina.
Original languageEnglish
Title of host publicationImage Processing
Subtitle of host publicationSPIE Medical Imaging, 2019, San Diego, California, United States
EditorsElsa D. Angelini, Bennett A. Landman
Place of PublicationBellingham
PublisherSPIE
Number of pages7
ISBN (Electronic)9781510625457
DOIs
Publication statusPublished - 1 Mar 2019
EventSPIE Medical Imaging 2019 - San Diego, United States
Duration: 16 Feb 201921 Feb 2019
http://spie.org/MI/entireprogram/2019-2-20?print=2&SSO=1

Publication series

NameProceedings of SPIE
Volume10949
ISSN (Electronic)1605-7422

Conference

ConferenceSPIE Medical Imaging 2019
CountryUnited States
CitySan Diego
Period16/02/1921/02/19
Internet address

Fingerprint

Biomarkers
Image analysis
Pipelines
Neural networks
Medical problems
Feature extraction
Screening
Chemical activation
Processing

Keywords

  • Deep Learning
  • Distillation
  • Maastricht Study
  • Retinal Biomarkers
  • Retinal Image Analysis

Cite this

Heslinga, F., Pluim, J., Dasht Bozorg, B., Berendschot, T., Houben, A. J. H. M., Henry, R. M. A., & Veta, M. (2019). Approximation of a pipeline of unsupervised retina image analysis methods with a CNN. In E. D. Angelini, & B. A. Landman (Eds.), Image Processing: SPIE Medical Imaging, 2019, San Diego, California, United States [109491N] (Proceedings of SPIE; Vol. 10949). Bellingham: SPIE. https://doi.org/10.1117/12.2512393
Heslinga, Friso ; Pluim, Josien ; Dasht Bozorg, Behdad ; Berendschot, Tos ; Houben, Alfons J.H.M. ; Henry, Ronald M.A. ; Veta, Mitko. / Approximation of a pipeline of unsupervised retina image analysis methods with a CNN. Image Processing: SPIE Medical Imaging, 2019, San Diego, California, United States. editor / Elsa D. Angelini ; Bennett A. Landman. Bellingham : SPIE, 2019. (Proceedings of SPIE).
@inproceedings{8b7e9c0257934c1e90e798c9a4aa9e98,
title = "Approximation of a pipeline of unsupervised retina image analysis methods with a CNN",
abstract = "A pipeline of unsupervised image analysis methods for extraction of geometrical features from retinal fundus images has previously been developed. Features related to vessel caliber, tortuosity and bifurcations, have been identified as potential biomarkers for a variety of diseases, including diabetes and Alzheimer's. The current computationally expensive pipeline takes 24 minutes to process a single image, which impedes implementation in a screening setting. In this work, we approximate the pipeline with a convolutional neural network (CNN) that enables processing of a single image in a few seconds. As an additional benefit, the trained CNN is sensitive to key structures in the retina and can be used as a pretrained network for related disease classification tasks. Our model is based on the ResNet-50 architecture and outputs four biomarkers that describe global properties of the vascular tree in retinal fundus images. Intraclass correlation coefficients between the predictions of the CNN and the results of the pipeline showed strong agreement (0.86 - 0.91) for three of four biomarkers and moderate agreement (0.42) for one biomarker. Class activation maps were created to illustrate the attention of the network. The maps show qualitatively that the activations of the network overlap with the biomarkers of interest, and that the network is able to distinguish venules from arterioles. Moreover, local high and low tortuous regions are clearly identified, confirming that a CNN is sensitive to key structures in the retina.",
keywords = "Deep Learning, Distillation, Maastricht Study, Retinal Biomarkers, Retinal Image Analysis",
author = "Friso Heslinga and Josien Pluim and {Dasht Bozorg}, Behdad and Tos Berendschot and Houben, {Alfons J.H.M.} and Henry, {Ronald M.A.} and Mitko Veta",
year = "2019",
month = "3",
day = "1",
doi = "10.1117/12.2512393",
language = "English",
series = "Proceedings of SPIE",
publisher = "SPIE",
editor = "Angelini, {Elsa D.} and Landman, {Bennett A.}",
booktitle = "Image Processing",
address = "United States",

}

Heslinga, F, Pluim, J, Dasht Bozorg, B, Berendschot, T, Houben, AJHM, Henry, RMA & Veta, M 2019, Approximation of a pipeline of unsupervised retina image analysis methods with a CNN. in ED Angelini & BA Landman (eds), Image Processing: SPIE Medical Imaging, 2019, San Diego, California, United States., 109491N, Proceedings of SPIE, vol. 10949, SPIE, Bellingham, SPIE Medical Imaging 2019, San Diego, United States, 16/02/19. https://doi.org/10.1117/12.2512393

Approximation of a pipeline of unsupervised retina image analysis methods with a CNN. / Heslinga, Friso; Pluim, Josien; Dasht Bozorg, Behdad; Berendschot, Tos; Houben, Alfons J.H.M.; Henry, Ronald M.A.; Veta, Mitko.

Image Processing: SPIE Medical Imaging, 2019, San Diego, California, United States. ed. / Elsa D. Angelini; Bennett A. Landman. Bellingham : SPIE, 2019. 109491N (Proceedings of SPIE; Vol. 10949).

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

TY - GEN

T1 - Approximation of a pipeline of unsupervised retina image analysis methods with a CNN

AU - Heslinga, Friso

AU - Pluim, Josien

AU - Dasht Bozorg, Behdad

AU - Berendschot, Tos

AU - Houben, Alfons J.H.M.

AU - Henry, Ronald M.A.

AU - Veta, Mitko

PY - 2019/3/1

Y1 - 2019/3/1

N2 - A pipeline of unsupervised image analysis methods for extraction of geometrical features from retinal fundus images has previously been developed. Features related to vessel caliber, tortuosity and bifurcations, have been identified as potential biomarkers for a variety of diseases, including diabetes and Alzheimer's. The current computationally expensive pipeline takes 24 minutes to process a single image, which impedes implementation in a screening setting. In this work, we approximate the pipeline with a convolutional neural network (CNN) that enables processing of a single image in a few seconds. As an additional benefit, the trained CNN is sensitive to key structures in the retina and can be used as a pretrained network for related disease classification tasks. Our model is based on the ResNet-50 architecture and outputs four biomarkers that describe global properties of the vascular tree in retinal fundus images. Intraclass correlation coefficients between the predictions of the CNN and the results of the pipeline showed strong agreement (0.86 - 0.91) for three of four biomarkers and moderate agreement (0.42) for one biomarker. Class activation maps were created to illustrate the attention of the network. The maps show qualitatively that the activations of the network overlap with the biomarkers of interest, and that the network is able to distinguish venules from arterioles. Moreover, local high and low tortuous regions are clearly identified, confirming that a CNN is sensitive to key structures in the retina.

AB - A pipeline of unsupervised image analysis methods for extraction of geometrical features from retinal fundus images has previously been developed. Features related to vessel caliber, tortuosity and bifurcations, have been identified as potential biomarkers for a variety of diseases, including diabetes and Alzheimer's. The current computationally expensive pipeline takes 24 minutes to process a single image, which impedes implementation in a screening setting. In this work, we approximate the pipeline with a convolutional neural network (CNN) that enables processing of a single image in a few seconds. As an additional benefit, the trained CNN is sensitive to key structures in the retina and can be used as a pretrained network for related disease classification tasks. Our model is based on the ResNet-50 architecture and outputs four biomarkers that describe global properties of the vascular tree in retinal fundus images. Intraclass correlation coefficients between the predictions of the CNN and the results of the pipeline showed strong agreement (0.86 - 0.91) for three of four biomarkers and moderate agreement (0.42) for one biomarker. Class activation maps were created to illustrate the attention of the network. The maps show qualitatively that the activations of the network overlap with the biomarkers of interest, and that the network is able to distinguish venules from arterioles. Moreover, local high and low tortuous regions are clearly identified, confirming that a CNN is sensitive to key structures in the retina.

KW - Deep Learning

KW - Distillation

KW - Maastricht Study

KW - Retinal Biomarkers

KW - Retinal Image Analysis

UR - http://www.scopus.com/inward/record.url?scp=85068327026&partnerID=8YFLogxK

U2 - 10.1117/12.2512393

DO - 10.1117/12.2512393

M3 - Conference contribution

T3 - Proceedings of SPIE

BT - Image Processing

A2 - Angelini, Elsa D.

A2 - Landman, Bennett A.

PB - SPIE

CY - Bellingham

ER -

Heslinga F, Pluim J, Dasht Bozorg B, Berendschot T, Houben AJHM, Henry RMA et al. Approximation of a pipeline of unsupervised retina image analysis methods with a CNN. In Angelini ED, Landman BA, editors, Image Processing: SPIE Medical Imaging, 2019, San Diego, California, United States. Bellingham: SPIE. 2019. 109491N. (Proceedings of SPIE). https://doi.org/10.1117/12.2512393