PaperlayerView exact revision on arXiv

Source-linked AI summary

DUNet: A deformable network for retinal vessel segmentation

Qiangguo Jin, Zhaopeng Meng, Tuan D. Pham, Qi Chen, Leyi Wei, Ran Su

arXiv:1811.01206v1cs.CV

TL;DR

Retinal vessel segmentation supports disease diagnosis but must handle complicated vessel structures and varied shapes. The paper proposes DUNet, a U-Net-based end-to-end network with deformable convolution and adaptive receptive fields. DUNet achieves strong performance across public datasets and preserves weak, tied, and detailed vessels.

  • Problem

    Retinal vessel segmentation is important for disease diagnosis, but vessel structures are complicated and vary in shape, scale, and orientation.

  • Method

    DUNet combines a U-shaped encoder-decoder with deformable convolution blocks, upsampling, and feature-map fusion for adaptive vessel segmentation.

  • Results

    DUNet achieves the highest accuracy among the compared models, reaching global accuracy of 0.9697 on DRIVE, 0.9729 on STARE, and 0.9724 on CHASE.

  • Takeaways & Limitations

    DUNet preserves more detailed, weak, and tied vessels than the compared networks in reported segmentation results.

Abstract

from arXiv · show

Automatic segmentation of retinal vessels in fundus images plays an important role in the diagnosis of some diseases such as diabetes and hypertension. In this paper, we propose Deformable U-Net (DUNet), which exploits the retinal vessels' local features with a U-shape architecture, in an end to end manner for retinal vessel segmentation. Inspired by the recently introduced deformable convolutional networks, we integrate the deformable convolution into the proposed network. The DUNet, with upsampling operators to increase the output resolution, is designed to extract context information and enable precise localization by combining low-level feature maps with high-level ones. Furthermore, DUNet captures the retinal vessels at various shapes and scales by adaptively adjusting the receptive fields according to vessels' scales and shapes. Three public datasets DRIVE, STARE and CHASE_DB1 are used to train and test our model. Detailed comparisons between the proposed network and the deformable neural network, U-Net are provided in our study. Results show that more detailed vessels are extracted by DUNet and it exhibits state-of-the-art performance for retinal vessel segmentation with a global accuracy of 0.9697/0.9722/0.9724 and AUC of 0.9856/0.9868/0.9863 on DRIVE, STARE and CHASE_DB1 respectively. Moreover, to show the generalization ability of the DUNet, we used another two retinal vessel data sets, one is named WIDE and the other is a synthetic data set with diverse styles, named SYNTHE, to qualitatively and quantitatively analyzed and compared with other methods. Results indicates that DUNet outperforms other state-of-the-arts.

I. INTRODUCTION

Retinal vessel segmentation matters for early diagnosis because vessel morphology can indicate disease, but complicated vessel structures make automated extraction difficult. The paper proposes DUNet to address these challenges with deformable, U-shaped segmentation.

  • Retinal vessel morphology can indicate diabetes, hypertension, and other pathological diseases, motivating accurate vessel extraction from fundus images.
  • Complicated vessel structures, high tortuosity, and varied shapes make retinal vessel segmentation challenging.
  • Manual segmentation is time-consuming and requires skilled staff, creating demand for automated retinal vessel segmentation.
  • DUNet uses an end-to-end U-shaped architecture with deformable convolution blocks to capture context, localize vessels, and adapt receptive fields to vessel geometry.
  • Prior vessel-segmentation methods include unsupervised morphology-based, curvature-based, multiscale, thresholding, and post-processing approaches.

B. Supervised method

Supervised retinal vessel segmentation learns mappings from hand-labeled images, but traditional systems depend heavily on manually designed features. Deep learning offers automated feature learning for more effective segmentation frameworks.

  • Supervised methods require hand-labeled data to map inputs to corresponding output labels.
  • Traditional supervised methods combine feature extractors with classifiers such as random forests, neural networks, SVMs, and AdaBoost.
  • Empirically defined features can require human intervention and introduce bias, motivating automated feature extraction.
  • Deep learning replaces handcrafted features with features learned automatically through hierarchical representations.
  • Deep learning vessel-segmentation studies have used CNNs, random forests, HED, and ConvNet ensembles for fundus-image processing.
  • The paper develops a fully convolutional framework for effective and automatic retinal vessel segmentation.

III. METHODOLOGY

The methodology introduces DUNet for retinal vessel segmentation and evaluates it on DRIVE, STARE, and CHASE. The workflow preprocesses and patches images before inference outputs are recomposed into complete maps.

  • DUNet integrates deformable convolution with U-Net to segment retinal vessels in fundus images.
  • Raw images are preprocessed and cropped into patches to create training and validation data for the compared networks.
  • DUNet and U-Net use 48 × 48 patches, whereas Deformable-ConvNet uses 29 × 29 patches for vessel classification.
  • The model is evaluated on DRIVE, STARE, and CHASE DB1, which provide fundus photographs with expert vascular annotations.

B. Image preprocessing and dataset preparation

The preparation pipeline enhances vessel-background contrast, trains on randomly sampled patches, and uses a deformable U-Net architecture to model vessels across shapes and scales.

  • B. Image preprocessing and dataset preparation: RGB fundus images are converted to single-channel images, normalized, enhanced with CLAHE, and processed with gamma correction.
  • B. Image preprocessing and dataset preparation: Randomly extracted 48 × 48 patches reduce computation while preserving surrounding local features for DUNet and U-Net training.
  • B. Image preprocessing and dataset preparation: Datasets are divided into training, validation, and test sets, with training patches sampled from the available images.
  • C. Deformable U-Net (DUNet): DUNet replaces standard convolution with deformable convolutional blocks in a U-shaped encoder-decoder architecture.
  • C. Deformable U-Net (DUNet): Deformable blocks learn local, dense, adaptive receptive fields to model retinal vessels with different shapes and scales.
  • C. Deformable U-Net (DUNet): Deformable convolution offsets sampling locations according to preceding feature maps, allowing adaptation to scale, shape, and orientation.

1) U-Net as the basic architecture:

The U-Net-style architecture combines encoding, decoding, upsampling, and skip connections for precise localization, while deformable convolutions adapt sampling to vessels’ varied shapes and scales.

  • U-Net architecture: U-Net uses symmetric encoding and decoding, with upsampling and merged feature maps enabling precise localization.The encoder produces lower-dimensional representations with richer filters, while the decoder reverses this process.
  • Deformable convolution: DUNet addresses vessels’ varied shapes and scales through deformable receptive fields that adapt to input features.This approach is motivated by the limitations of traditional multiscale or multidirectional filters.
  • Deformable convolution: Deformable convolution learns offsets from preceding feature maps, adapting sampling to scale, shape, and orientation.Offsets are added to the regular convolution grid and learned through additional convolutional layers.
  • Deformable convolution: The deformable-convolution illustration uses a 5 × 5 kernel whose offsets and convolution kernels are learned jointly.The offset field is generated from input patches and features, with 2N channels representing N 2D offsets.
  • Deformable convolution: Noninteger offsets require bilinear interpolation, adding computation compared with regular U-Net for more local and adaptive processing.The interpolation determines sampled values after the grid locations are shifted.

D. Compare with U-Net and Deformable-ConvNet

The study compares DUNet with U-Net and Deformable-ConvNet, which treats vessel segmentation as pixel classification using a local image patch.

  • Model comparisons: DUNet is compared against standard U-Net and Deformable-ConvNet as reference models.Deformable-ConvNet was originally introduced to classify whether individual pixels belong to vessels.
  • Deformable-ConvNet: Deformable-ConvNet classifies each selected pixel from neighborhood information in a centered 29×29 patch.The patch size was chosen to reduce calculation complexity while capturing local features.

E. Performance evaluation metrics

Performance is assessed with pixel-classification, discrimination, overlap, and balance metrics, including ACC, PPV, TPR, TNR, AUC, F1, and Jaccard similarity.

  • Core metrics: The evaluation uses Accuracy, Positive Predictive Value, True Positive Rate, True Negative Rate, and ROC Area Under Curve.These metrics measure overall correctness, precision, sensitivity, specificity, and ROC discrimination, respectively.
  • Metric definitions: TPR measures correctly identified positives, while TNR measures correctly identified negatives.TPR is also called sensitivity, and TNR is also called specificity.
  • Metric definitions: ACC is computed from true positives and true negatives over all positive and negative prediction outcomes.The formulation includes TP, TN, FP, and FN.
  • Overlap metrics: F-measure and Jaccard similarity additionally compare similarity and diversity between testing results and reference segmentations.Ground truth is denoted GT, while the segmentation result is denoted SR.

IV. EXPERIMENTAL RESULT

Experiments compare DUNet, U-Net, and Deformable-ConvNet across DRIVE, STARE, and CHASE using validation and test evaluations. DUNet generally achieves the strongest reported performance, including the highest accuracy across the three models on the test datasets and the largest ROC areas.

  • Validation results: DUNet achieved validation accuracy of 0.9650 and loss of 0.0919 on DRIVE, and validation accuracy of 0.9704 and loss of 0.0833 on CHASE.It had the highest validation accuracy and lowest loss on both datasets; on STARE, it had the second-highest accuracy and lowest loss.
  • Test evaluation: The test comparison evaluates PPV, TPR, TNR, ACC, F1, JS, and AUC across DRIVE, STARE, and CHASE.The three models are trained from scratch and compared using separate training, validation, and test splits.
  • Test results: 0.9697/0.9729/0.9724 global accuracy was reported for DUNet on DRIVE/STARE/CHASE, respectively.The corresponding model order in the comparison is Deformable-ConvNet, U-Net, and DUNet.
  • ROC analysis: DUNet’s ROC curves are closest to the top-left and have the largest AUC among the three models.The ROC comparisons are reported for DRIVE, STARE, and CHASE.

B. Retinal vessel segmentation results

DUNet preserves more retinal vessel detail than U-Net and Deformable-ConvNet, particularly for weak, tiny, tied, and closely spaced vessels in complex vascular structures.

  • DUNet detects weak or tied vessels that U-Net and Deformable-ConvNet may miss, preserving more segmentation detail.
  • In vascular junctions, DUNet segments tied vessels more successfully than the comparison models.Deformable-ConvNet extracts more detail than U-Net in some junction regions.
  • DUNet produces desirable segmentation results for tiny and weak vessels, where U-Net has difficulty and Deformable-ConvNet recovers some details.
  • The magnified comparisons cover vascular-junction and tiny-vessel patches from DRIVE, STARE, and CHASE.

C. Comparison against existing methods

DUNet is compared with standard and deep-learning segmentation methods across established and additional datasets, showing strong benchmark performance and competitive generalization.

  • DUNet achieves the highest global accuracy of 0.9697/0.9724 and AUC of 0.9856/0.9863 on DRIVE and CHASE among the compared methods.The comparison includes standard segmentation algorithms and deep neural network methods.
  • On STARE, DUNet does not outperform Liskowski et al. or Alom et al., but uses fewer training patch samples while reaching desirable results.
  • Using weights trained on DRIVE, DUNet produces qualitatively competitive results on WIDE and SYNTHE.WIDE contains 15 high-resolution wide-field RGB images and lacks ground-truth vessel-segmentation labels.
  • Quantitative evaluation on SYNTHE shows DUNet achieves the best overall performance among the three tested models.The SYNTHE images combine four distinct styles and use DRIVE-trained weights.

V. CONCLUSION

DUNet extends U-Net with deformable convolution blocks to segment retinal vessels pixel-wise while adapting receptive fields to vessel shape and scale. The authors report more detailed vessel extraction and state-of-the-art segmentation performance, while identifying the need for more data to validate the model.

  • DUNet replaces U-Net convolutional layers with deformable convolution blocks in a symmetric U-shape network for pixel-wise retinal vessel segmentation.Its encoder captures context, while the decoder combines low- and high-level features for localization.
  • Deformable convolution adapts receptive fields to retinal vessels’ shapes and scales by adding offsets to standard convolution sampling grids.
  • DUNet extracts more detailed vessels and exhibits state-of-the-art performance in retinal vessel segmentation.
  • The authors plan to incorporate more retinal vessel data to validate the proposed end-to-end model.They also plan to extend DUNet to three dimensions for medical image analysis tasks.
Loading 1811.01206v1…