Deep ensemble learning for accurate retinal vessel segmentation

https://doi.org/10.1016/j.compbiomed.2023.106829Get rights and content

Highlights

  • In this work, we propose a deep ensemble learning framework for the retinal vessel segmentation.

  • Our models outperform the existing approaches on several datasets, demonstrating that our models are more effective, superior, and robust for the retinal vessel segmentation.

  • Our model is capable of capturing the discriminative feature representations through introducing the ensemble strategy to integrate different base deep learning models.

Abstract

Significant progress has been made in deep learning-based retinal vessel segmentation in recent years. However, the current methods suffer from low performance and the robust of the models is not that good. Our work introduces an novel framework for retinal vessel segmentation based on deep ensemble learning. The results of benchmarking comparisons indicate that our model outperforms the existing ones on multiple datasets, demonstrating that our models are more effective, superior, and robust for the retinal vessel segmentation. It evinces the capability of our model to capture the discriminative feature representations through introducing the ensemble strategy to integrate different base deep learning models like pyramid vision Transformer and FCN-Transformer. We expect our proposed method can benefit and accelerate the development of accurate retinal vessel segmentation in this field.

Introduction

Cardiovascular disease remains the predominant cause of mortality on a global scale, and it will continue to rise in the next 10 years [[1], [2], [3]]. The fundus retinal blood vessels are unique in that they are the only blood vessels in the body that can be directly observed. Studies have found that their morphological and functional changes are associated over the course of its development of many cardiovascular diseases, including coronary heart disease, atherosclerosis, and hypertension, and have the potential to become a tool for the assessment and prediction of related diseases [[4], [5], [6], [7]]. Therefore, research on automated retinal analysis is undoubted of great significance to public health.

Earlier, researches on Vessel segmentation were mainly based on traditional handcrafted features. For instance, quadrature filters [8], matched filters [9], and Gabor filters [10,11]. Specifically, the alternative incorporates all vessel widths and orientations with a predefined kernel bank. Handcrafted features, in fact, have inevitable limitations like our capability to analytically model the segmentation process [12,13]. There are also some methods put forward by Kaur [14], including region growing [15], piece-wise thresholding [16], and concavity measurements [17]. To trace the vascular structure, researchers have used some graph-theoretic techniques, apart from handcrafted features. For example, the fast marching algorithm [18] and shortest-path tracking [19]. AlexNet won the championship of ILSVRC in 2012, marking a new breakthrough for convolutional neural networks (CNNs) on ImageNet dataset. Since then, deep learning has developed rapidly, making new breakthroughs in various fields [[20], [21], [22], [23], [24], [25], [26]]. At present, U-NET is the most successful deep network used for blood vessel segmentation. It adds feature upsampling on the basis of traditional convolutional layer and fuses the location information with the semantic information through skip connections. As mentioned above, prior deep learning approaches applied a classifier to a series of small patches that were divided from an image. Actually, that's exactly what Dasgupta and Singh did [27]. Numerous adaptations to the initial U-net design have been developed. Alom et al. added the addition of residual connections and the application of a recursive training strategy to the U-net, which improves performance but increases training time. To learn more features, Zhuang et al. introduced a multi-path network that consists of two U-net architectures were serially combined [28]. However, it leads to greatly increased training time. Besides, Jiang and Tan incorporated the U-net architecture augmented with conditional generative adversarial networks (GANs) resulted in improved performance and get a more balanced result, Jin et al. incorporating deformable CNNs resulted in better vascular feature construction [29], but they both significantly increased complexity.

To improve the segmentation performance, Our work introduces an novel framework for retinal vessel segmentation based on deep ensemble learning. Benchmarking comparison results show that our model surpasses existing approaches on several datasets, demonstrating that our models are more effective, superior, and robust for the retinal vessel segmentation. It evinces the capability of our model to capture the discriminative feature representations through introducing the ensemble strategy to integrate different base deep learning models like pyramid vision Transformer and FCN-Transformer. We expect our proposed method can benefit and accelerate the development of accurate retinal vessel segmentation in this field.

Section snippets

Methodology

We establish the overall framework illustrated in Fig. 1. First, we obtain the retinal vessel images from the three datasets. Then we used data preprocessing includes image cropping and green channel extraction from original retinal images. And we used the common augmented operations including image flipping and so on. After this, we used three convolutional vision transformer base networks to as the independent base model predicted the results. It is note that the independent base networks

Comparison with existing methods

To strengthen the evidence supporting the effectiveness of our proposed approach, we conducted a comparative analysis with the most advanced models currently available. These models included U-net architecture and its extensions, including LadderNet [38], R2Unet [37], and DUNet [29]. A summary of the performance of the compared methods on various datasets is presented in Table 1, Table 2, Table 3. It is worth noting that most of the current state-of-the-art methods exhibit lower F1 scores

Conclusion

The contribution of this work is a deep ensemble learning framework that is designed specifically for retinal vessel segmentation. Benchmarking comparison results show that our models perform better that the existing approaches on several datasets, demonstrating that our models are more effective, superior, and robust for the retinal vessel segmentation. Via feature analysis, we found that our model is capable of capturing the discriminative feature representations through introducing the

Funding

This study was supported by Scientific Research Innovation Fund of the First Hospital of Harbin Medical University (2017B003), the Natural Science Foundation of Heilongjiang Province (Grant No. LH2020H061), and the Foundation of Academy of Medical Sciences of Heilongjiang Province (Grant No. 201702).

Declaration of cCompeting iInterest

There is no competing financial interest to declare.

References (41)

  • G. Läthén et al.

    Blood vessel segmentation using multi-scale quadrature filtering

    Pattern Recogn. Lett.

    (2010)
  • X. Zeng

    Toward better drug discovery with knowledge graph

    Curr. Opin. Struct. Biol.

    (2022)
  • X. Zeng

    Deep generative molecular design reshapes drug discovery

    Cell Rep. Med.

    (2022)
  • Q. Jin

    DUNet: a deformable network for retinal vessel segmentation

    Knowl. Base Syst.

    (2019)
  • W. Zengwu et al.

    An essential introduction to the annual report on cardiovascular health and diseases in China (2021)

    Chinese General Practice

    (2022)
  • Y. Hu et al.

    Integration of multiple-omics data to analyze the population-specific differences for coronary artery disease

    Comput. Math. Methods Med.

    (2021)
  • X. Zeng

    Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework

    Nat. Mach. Intell.

    (2022)
  • M.K. Ikram

    Retinal vascular caliber measurements: clinical significance, current knowledge and future perspectives

    Ophthalmologica

    (2013)
  • Y. Hu

    rs1990622 variant associates with Alzheimer's disease and regulates TMEM106B expression in human brain tissues

    BMC Med.

    (2021)
  • Y. Hu

    rs34331204 regulates TSPAN13 expression and contributes to Alzheimer's disease with sex differences

    Brain

    (2020)
  • Z. Li

    DSN-DDI: an accurate and generalized framework for drug–drug interaction prediction by dual-view representation learning

    Briefings Bioinf.

    (2023)
  • S. Chaudhuri

    Detection of blood vessels in retinal images using two-dimensional matched filters

    IEEE Trans. Med. Imag.

    (1989)
  • F. Farokhian et al.

    Blood vessels detection and segmentation in retina using Gabor filters

  • Z. Yavuz et al.

    Retinal blood vessel segmentation using Gabor filter and top-hat transform

  • Y. Hu

    Mendelian Randomization Highlights Causal Association between Genetically Increased C-Reactive Protein Levels and Reduced Alzheimer's Disease Risk

    (2022)
  • Y. Hu

    Cognitive performance protects against Alzheimer’s disease independently of educational attainment and intelligence

    Mol. Psychiatr.

    (2022)
  • D. Kaur et al.

    M. Computing, Various image segmentation techniques: a review

    Int. J. Comput. Sci. Mobile Comput.

    (2014)
  • M.E. Martínez-Pérez

    Retinal blood vessel segmentation by means of scale-space analysis and region growing

  • A. Hoover et al.

    Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response

    IEEE Trans. Med. Imag.

    (2000)
  • B.S. Lam et al.

    General retinal vessel segmentation using regularization-based multiconcavity modeling

    IEEE Trans. Med. Imag.

    (2010)
  • Cited by (8)

    View all citing articles on Scopus
    View full text