Deep Retinal Image Segmentation With Regularization Under Geometric Priors

Venkateswararao Cherukuri, Vijay Kumar B. G, Raja Bala, Vishal MongaIEEE TIP 2020 · DOI

TL;DR

This paper introduces DRIS-GP, a novel deep learning framework for retinal vessel segmentation that incorporates geometric priors to improve accuracy and efficiency. By learning domain-specific features and regularizing filters, DRIS-GP achieves state-of-the-art results while requiring fewer parameters and less computation than existing methods.

What Problem Does This Paper Address?

Accurate segmentation of blood vessels in retinal images is crucial for diagnosing eye diseases like diabetic retinopathy and glaucoma. However, this task is challenging due to low image contrast, varying vessel sizes and thicknesses, and the presence of interfering pathologies like micro-aneurysms. While deep learning has significantly improved segmentation accuracy, existing methods often rely heavily on data and can be computationally expensive. This paper aims to address these limitations by developing a more efficient and accurate segmentation method that leverages prior knowledge about retinal vessel geometry.

Methods

The core of the proposed method, DRIS-GP, is a two-component network:

  1. Representation Network: This network learns geometric features specific to retinal images. It's designed to be "domain-enriched" by incorporating two novel regularization terms into the learning process:
  2. Orientation Diversity Constraint: Encourages filters to capture a wide range of vessel orientations (0-180 degrees).
  3. Data-Adaptive Noise Constraint: Penalizes filters that respond to image patches containing noise or structures that resemble vessels but are not, thus improving robustness to false positives.
  4. Multi-Scale Extension: To handle varying vessel thicknesses, the representation layer is extended to operate at multiple spatial scales.
  5. Task Network: A computationally efficient residual network (ResNet-inspired) is used for the actual pixel-level segmentation. This network leverages the rich, geometrically informed features from the representation layer. The authors simplify the standard ResNet block by removing Batch Normalization layers and using only a single convolutional layer, enabled by the strong features from the preceding layer.

Both the representation and task networks are jointly learned.

Datasets

The authors evaluate their method on three widely used benchmark datasets for retinal vessel segmentation:

  • DRIVE: Contains 40 fundus images with manually labeled ground truth.
  • STARE: Contains 20 manually labeled fundus images.
  • CHASEDB1: Contains 28 manually labeled fundus images.

They also perform experiments on the HRF (High-Resolution Fundus Images) dataset, which has a higher resolution and different image characteristics.

Key Insights and Ideas

  • Geometric Priors Enhance Deep Learning: The paper demonstrates that incorporating domain-specific geometric knowledge (like vessel orientation and noise characteristics) into deep learning filters can significantly improve performance and efficiency, moving beyond purely data-driven approaches.
  • Regularization for Robustness: The proposed orientation diversity and data-adaptive noise regularizers are crucial for learning filters that are both sensitive to vessel structures and robust to noise and false positives.
  • Simplified Network Architecture: By learning highly informative features in the representation layer, the subsequent task network can be significantly simplified (e.g., fewer layers, no Batch Normalization), leading to faster inference and fewer parameters.
  • Multi-Scale Processing is Key for Thin Vessels: The multi-scale extension of the representation layer is particularly effective in accurately segmenting thin vessels, a common challenge in the field.

Key Results

  • State-of-the-Art Performance: MS-DRIS-GP consistently achieved state-of-the-art results across all three benchmark datasets (DRIVE, STARE, CHASEDB1) in terms of F1-score, Accuracy, Sensitivity, and Specificity. For example, on DRIVE, it achieved an F1 score of 0.8220.
  • Improved Thin Vessel Segmentation: The method showed superior performance in detecting thin vessels compared to other specialized methods like ML-UNET and JL-UNET, achieving higher AUC and specificity for thin vessels.
  • Computational Efficiency: MS-DRIS-GP requires significantly fewer parameters (around 1.3 x 10^5) and has a much faster inference time (26ms on NVIDIA Titan X GPU) compared to other deep learning methods (e.g., PDSN at 92s, ML-UNET at 33ms).
  • Robustness to Limited Data: The method exhibits graceful degradation with fewer training samples and outperforms competing methods significantly at 10% training data, highlighting the benefit of domain priors.

Limitations and Future Work

  • Limitations: While the paper demonstrates strong performance, the authors acknowledge that the current approach focuses on geometric priors. They suggest that incorporating higher-order structural priors on the network outputs could further enhance segmentation.
  • Future Work: The authors propose exploring additional ways to integrate domain knowledge into deep retinal segmentation networks. Specific ideas include incorporating structural priors on the network outputs to capture higher-order vessel structures. They also suggest applying these ideas to other curvilinear segmentation problems like lung vessel detection, facial wrinkle detection, and road detection.

Summary updated

BibTeX
@article{cherukuri2020deep,
  author = {Venkateswararao Cherukuri and Vijay Kumar B. G and Raja Bala and Vishal Monga},
  title = {{Deep Retinal Image Segmentation With Regularization Under Geometric Priors}},
  journal = {IEEE TIP},
  year = {2020},
  doi = {10.1109/tip.2019.2946078},
}

Loading related content...

Figure