Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Better feature extraction using multi-encoder convolutional neural networks for optic cup segmentation from digital fundus images

Better feature extraction using multi-encoder convolutional neural networks for optic cup... PurposeGlaucoma is an eye disease that is chronic, asymptomatic, and cannot be cured once it progresses. An important step in clinical analysis of glaucoma is to measure the cup-to-disc ratio (CDR). Optic cup segmentation is a challenging task (as compared to detecting the optic disk, for instance), due to poor contrast on the cup boundary region, and occlusion from veins and arteries. Contemporary systems are based on image processing/computer vision and/or machine learning. However, obtaining accurate optic cup segmentation over large datasets is still a challenge.MethodsWe propose a novel asymmetric “multi-encoder U-Net”/Y-Net architecture with Inception and context blocks in the bottleneck layer. The architecture has an ResNet34-based primary encoder and a light-and-efficient EfficientNetB0 auxiliary encoder. The asymmetry involves avoiding multi-stage skip connections from the auxiliary encoder to the decoder. This avoids the complexity of feature map concatenation at different levels. The Inception block in the bottleneck layer performs feature enrichment. Different receptive fields in parallel paths result in multi-scale optic cup features. The next cascaded context block helps maintain spatial consistency of the multi-scale feature maps.Results and discussionWe have experimented extensively on four public datasets, and the challenging AIIMS community camp (private) dataset. The proposed network outperforms the state of the art with an average Dice coefficient of 91.11% and 87.77% on the Drishti-GS Sivaswamy et al. (2014) and Refugee Maninis and Pont-Tuset (2010) public datasets. Our ablation studies with different competing architectures also show the proposed method achieving the highest Dice coefficient and cup overlap percentage. The training itself achieves a much lower train-validation loss, as seen over a large number of epochs.ConclusionThe novel architecture has each sub-part geared towards getting good optic cup segmentation performance across a large number of datasets. The network shows robust segmentation performance on challenging images with various retinal artifacts (blurring, poor illumination, and clinical pathologies). http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Research on Biomedical Engineering Springer Journals

Better feature extraction using multi-encoder convolutional neural networks for optic cup segmentation from digital fundus images

Loading next page...
 
/lp/springer-journals/better-feature-extraction-using-multi-encoder-convolutional-neural-v1I4shqPes

References (33)

Publisher
Springer Journals
Copyright
Copyright © The Author(s), under exclusive licence to The Brazilian Society of Biomedical Engineering 2023. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
ISSN
2446-4732
eISSN
2446-4740
DOI
10.1007/s42600-022-00249-5
Publisher site
See Article on Publisher Site

Abstract

PurposeGlaucoma is an eye disease that is chronic, asymptomatic, and cannot be cured once it progresses. An important step in clinical analysis of glaucoma is to measure the cup-to-disc ratio (CDR). Optic cup segmentation is a challenging task (as compared to detecting the optic disk, for instance), due to poor contrast on the cup boundary region, and occlusion from veins and arteries. Contemporary systems are based on image processing/computer vision and/or machine learning. However, obtaining accurate optic cup segmentation over large datasets is still a challenge.MethodsWe propose a novel asymmetric “multi-encoder U-Net”/Y-Net architecture with Inception and context blocks in the bottleneck layer. The architecture has an ResNet34-based primary encoder and a light-and-efficient EfficientNetB0 auxiliary encoder. The asymmetry involves avoiding multi-stage skip connections from the auxiliary encoder to the decoder. This avoids the complexity of feature map concatenation at different levels. The Inception block in the bottleneck layer performs feature enrichment. Different receptive fields in parallel paths result in multi-scale optic cup features. The next cascaded context block helps maintain spatial consistency of the multi-scale feature maps.Results and discussionWe have experimented extensively on four public datasets, and the challenging AIIMS community camp (private) dataset. The proposed network outperforms the state of the art with an average Dice coefficient of 91.11% and 87.77% on the Drishti-GS Sivaswamy et al. (2014) and Refugee Maninis and Pont-Tuset (2010) public datasets. Our ablation studies with different competing architectures also show the proposed method achieving the highest Dice coefficient and cup overlap percentage. The training itself achieves a much lower train-validation loss, as seen over a large number of epochs.ConclusionThe novel architecture has each sub-part geared towards getting good optic cup segmentation performance across a large number of datasets. The network shows robust segmentation performance on challenging images with various retinal artifacts (blurring, poor illumination, and clinical pathologies).

Journal

Research on Biomedical EngineeringSpringer Journals

Published: Mar 1, 2023

Keywords: Retinal images; Optic cup; Deep convolutional neural network (DCNN); U-Net; Image segmentation

There are no references for this article.