PaperlayerView exact revision on arXiv

Source-linked AI summary

Patch-based Output Space Adversarial Learning for Joint Optic Disc and Cup Segmentation

Shujun Wang, Lequan Yu, Xin Yang, Chi-Wing Fu, Pheng-Ann Heng

arXiv:1902.07519v1cs.CV

TL;DR

Cross-dataset domain shift limits joint optic-disc and optic-cup segmentation, despite strong within-dataset deep-learning performance. The paper proposes pOSAL, combining a lightweight morphology-aware segmenter with patch-based output-space adversarial adaptation using unlabeled target images. It reports improved segmentation across three public datasets and first place in the MICCAI 2018 challenge.

  • Problem

    Domain shift between retinal fundus datasets hinders deep networks from generalizing across scanners and institutions, limiting robust OD and OC segmentation for glaucoma assessment.

  • Method

    pOSAL combines a lightweight segmentation network and morphology-aware loss with unsupervised patch-based output-space adversarial learning between labeled source and unlabeled target images.

  • Results

    The framework improves segmentation on Drishti-GS, RIM-ONE-r3, and REFUGE and achieves first place in the MICCAI 2018 OD and OC segmentation tasks.

  • Takeaways & Limitations

    Patch-based output-space adaptation captures local output statistics and guides similar source- and target-domain predictions for joint OD and OC segmentation.

  • Takeaways & Limitations

    Training requires unlabeled target-domain images, and a new network must be retrained when images come from a new target domain.

Abstract

from arXiv · show

Glaucoma is a leading cause of irreversible blindness. Accurate segmentation of the optic disc (OD) and cup (OC) from fundus images is beneficial to glaucoma screening and diagnosis. Recently, convolutional neural networks demonstrate promising progress in joint OD and OC segmentation. However, affected by the domain shift among different datasets, deep networks are severely hindered in generalizing across different scanners and institutions. In this paper, we present a novel patchbased Output Space Adversarial Learning framework (pOSAL) to jointly and robustly segment the OD and OC from different fundus image datasets. We first devise a lightweight and efficient segmentation network as a backbone. Considering the specific morphology of OD and OC, a novel morphology-aware segmentation loss is proposed to guide the network to generate accurate and smooth segmentation. Our pOSAL framework then exploits unsupervised domain adaptation to address the domain shift challenge by encouraging the segmentation in the target domain to be similar to the source ones. Since the whole-segmentationbased adversarial loss is insufficient to drive the network to capture segmentation details, we further design the pOSAL in a patch-based fashion to enable fine-grained discrimination on local segmentation details. We extensively evaluate our pOSAL framework and demonstrate its effectiveness in improving the segmentation performance on three public retinal fundus image datasets, i.e., Drishti-GS, RIM-ONE-r3, and REFUGE. Furthermore, our pOSAL framework achieved the first place in the OD and OC segmentation tasks in MICCAI 2018 Retinal Fundus Glaucoma Challenge.

I. INTRODUCTION

The paper addresses poor cross-dataset generalization in joint optic disc and cup segmentation caused by domain shift. It introduces pOSAL, combining morphology-aware segmentation with patch-based output-space adversarial adaptation using labeled source and unlabeled target images.

  • Motivation: Glaucoma causes irreversible vision loss, motivating objective optic disc and cup segmentation for scalable screening.Manual measurements are time-consuming, while cup-to-disc ratio provides a quantitative glaucoma-risk indicator.
  • Problem: Deep networks can segment optic structures effectively within a dataset but generalize poorly across datasets with different imaging conditions.Domain shift arises from differences in scanners, resolution ratios, light intensities, and parameter settings.
  • Proposed Approach: pOSAL combines a lightweight segmentation network, morphology-aware loss, and adversarial learning to produce smooth predictions consistent across source and target domains.The framework uses annotated source images and unannotated target images to reduce target-domain degradation.
  • Evaluation: The framework is evaluated on Drishti-GS, RIM-ONE-r3, and REFUGE, and achieved first place in both OD and OC segmentation at the MICCAI 2018 challenge.These experiments are presented as evidence of effectiveness across public retinal fundus datasets.
  • Prior Work and Gap: Existing adaptation strategies use target annotations, latent-feature alignment, or image transfer, but pOSAL instead aligns spatial and geometric structures in the output space.The paper identifies label-space structure as highly correlated across domains and uses patch-based output-space adaptation.

III. METHODOLOGY

The methodology uses a two-stage pipeline that first extracts optic-disc regions and then segments the optic disc and cup. Its segmentation network is based on DeepLabv3+ with a MobileNetV2 backbone.

  • Segmentation Pipeline: The cropped source and target regions are fed into the segmentation network to generate optic-disc and optic-cup predictions.The shared extraction process supplies corresponding disc-centered inputs for both domains.
  • ROI Extraction: An extraction network first segments the optic disc to locate and crop a 512 × 512 region of interest for subsequent segmentation.The network is trained on resized source images and source optic-disc labels, then applied to both source and target images.
  • Segmentation Network: The segmentation network follows DeepLabv3+ while replacing its Xception backbone with the lighter MobileNetV2.This design targets lower parameter and computation costs while retaining multi-scale segmentation capabilities.

B. Segmentation Network with Morphology-aware Loss

The segmentation network combines a lightweight DeepLabv3+-style architecture with a morphology-aware loss to jointly produce accurate, smooth OD and OC masks.

  • The network replaces DeepLabv3+’s Xception backbone with MobileNetV2 to reduce parameters and computation cost.
  • A 1×1 convolution with two Sigmoid-activated output channels simultaneously generates OD and OC probability maps.
  • The morphology-aware loss combines Dice overlap supervision with a smoothness term for OD and OC segmentation.
  • The Dice loss measures overlap between predicted probability maps and binary ground-truth masks.
  • The smoothness loss encourages neighboring pixels with the same ground-truth class to receive similar predicted probabilities.
  • The loss weights are λ1 = 0.4, λ2 = 0.6, and λ3 = 1.0, with greater emphasis placed on OC because its boundaries are less clear.

C. Patch-based Output Space Adversarial Learning

The framework adapts segmentation across domains in output space rather than feature or image space, using adversarial learning to align target masks with source-mask structure.

  • Feature-space adaptation is challenging because segmentation features must encode appearance, shape, context, and semantics, whereas output masks retain correlated spatial structure across domains.
  • A patch-level discriminator compares source and target segmentation-mask distributions, while the segmentation network learns to fool it with target predictions resembling source predictions.
  • The adversarial process preserves geometric structure in segmentation masks while addressing appearance shifts between domains.

1) Patch Discriminator:

The patch discriminator performs local output-space discrimination, encouraging target predictions to match source predictions at the level of overlapping mask patches.

  • 1) Patch Discriminator: PatchGAN classifies whether each overlapping m×n patch from a predicted mask matches the distribution of source-domain prediction patches.
  • 1) Patch Discriminator: Unlike image-level or pixel-level adversarial learning, PatchGAN captures local output-space statistics and focuses segmentation on local structural similarity.
  • 1) Patch Discriminator: The discriminator is a five-layer fully convolutional network producing a 16×16 map, where each output pixel corresponds to a 94×94 input probability-map patch.
  • 1) Patch Discriminator: The discriminator is trained to identify target-domain predictions, while the segmentation network uses adversarial loss to make them appear source-like.
  • 1) Patch Discriminator: Training alternates source and target images, combining source morphology-aware segmentation loss with target adversarial loss.

3) Training Strategy:

The framework is trained with source annotations and unlabeled target images across retinal datasets with different image appearances, enabling evaluation under domain shift.

  • 3) Training Strategy: Training feeds source and target images alternately and minimizes the segmentation-network and discriminator objectives during each iteration.
  • 3) Training Strategy: The experiments use Drishti-GS, RIM-ONE-r3, and REFUGE, assigning REFUGE training images as source and the remaining dataset partitions as targets.
  • 3) Training Strategy: Source and target images come from different cameras, producing differences in color and texture.
  • 3) Training Strategy: Table II compares pOSAL with pOSALseg-S on Drishti-GS and RIM-ONE-r3 testing datasets to assess output-space adversarial learning.
  • 3) Training Strategy: The implementation first trains the segmentation network with source images and annotations, then trains the complete pOSAL framework end-to-end through adversarial learning.

C. Evaluation Metrics

The study evaluates joint optic disc and cup segmentation with Dice coefficients and absolute cup-to-disc-ratio error, using separate unlabeled target images for adaptation and testing. Results show patch-based output-space adversarial learning improves segmentation across both target datasets.

  • Evaluation Metrics: Dice coefficients measure segmentation overlap, while δ measures the absolute difference between predicted and ground-truth vertical cup-to-disc ratios.The vertical CDR is clinically relevant for glaucoma screening.
  • Evaluation Metrics: Under domain adaptation, unlabeled target images used for training differ from target testing images to preserve a fair comparison.REFUGE training images served as the source domain, while Drishti-GS and RIM-ONE-r3 training images served as target-domain data.
  • Evaluation Metrics: pOSAL consistently improves optic cup and disc Dice coefficients and δ over pOSALseg-S on Drishti-GS and RIM-ONE-r3.The comparison isolates the effect of output-space adversarial learning.
  • Evaluation Metrics: 4.3% and 8.6% DI improvements occur for cup and disc segmentation on RIM-ONE-r3, versus 2.2% and 2.1% on Drishti-GS.The larger gains on RIM-ONE-r3 coincide with a larger domain discrepancy from REFUGE.
  • Evaluation Metrics: Qualitative Drishti-GS results show pOSAL preserves shape priors and produces more accurate OD and OC boundaries than pOSALseg-S.The baseline often localizes approximate regions but fails at low-contrast boundaries.

2) Qualitative Results:

The proposed segmentation network is compared with supervised methods on Drishti-GS and RIM-ONE-r3. Its morphology-aware design achieves stronger Dice performance, while pOSAL approaches supervised performance on Drishti-GS.

  • Comparison with other Segmentation Methods: The comparison uses REFUGE challenge testing results for OD and OC segmentation, with the top three items bolded for each metric.The table reports challenge-dataset outcomes rather than only the proposed model's internal ablation.
  • Comparison with other Segmentation Methods: pOSALseg-T produces better optic cup and disc Dice coefficients than other supervised methods on both Drishti-GS and RIM-ONE-r3.This evaluation uses dataset-specific training splits and testing images.
  • Comparison with other Segmentation Methods: pOSAL performance on Drishti-GS is very close to that of the supervised methods.This supports the effectiveness of patch-based output-space adversarial learning under domain adaptation.

4) Comparison with Different Domain Adaptation Approaches:

The paper compares pOSAL with domain-adaptation methods and evaluates its downstream glaucoma-screening performance. It also tests adaptation on REFUGE validation data and examines morphology-aware loss effects.

  • Comparison with Different Domain Adaptation Approaches: pOSAL is compared with Cycle-GAN-based, latent-feature-alignment, and other unsupervised domain-adaptation approaches because prior OD and OC segmentation work lacked domain-adaptation studies.The comparison spans methods developed for medical and natural image-processing tasks.
  • Performance of Glaucoma Screening: Vertical CDR values computed from segmented OD and OC masks are normalized across testing images for glaucoma-screening evaluation.ROC curves and AUC are reported for Drishti-GS and RIM-ONE-r3.
  • Results of the REFUGE Challenge: On the REFUGE challenge testing dataset, pOSAL ranks first overall and outperforms the second-ranking Masker team by around 1.4% on optic-disc DI.The challenge ranking uses a weighted sum of cup DI, disc DI, and CDR error ranks.
  • Results of the REFUGE Challenge: pOSAL improves optic cup and disc DI on the REFUGE validation dataset when compared with pOSALseg-S.Half of the 400 validation images are used as unlabeled target training data and half for testing.
  • Comparison with Different Domain Adaptation Approaches: Combining Dice and smooth losses yields the best OD and OC DI, indicating that the morphology-aware loss produces high-quality predictions.The loss comparison uses a split of 320 REFUGE training images for training and 80 for evaluation.
  • Performance of Glaucoma Screening: The glaucoma-screening AUC on the REFUGE challenge testing dataset is 0.9644, ranking third onsite.Ground-truth glaucoma labels were unavailable, so only AUC is reported.

V. DISCUSSION

The discussion connects clinically motivated OD/OC segmentation with a two-stage ROI pipeline and a computationally efficient backbone, while identifying dependence on unlabeled target-domain images as a practical boundary.

  • Clinical motivation: High-quality automatic OD/OC segmentation is sought because optic disc-to-cup measurements support glaucoma screening, but clinical deployment faces annotation and domain-shift gaps.The paper notes that clinical applications may involve missing, noisy, or sparse annotations and differences between training and real testing images.
  • ROI extraction: The two-stage pipeline improves segmentation by first extracting an OD-guided ROI, with an ROI roughly twice the OD size covering all OD and OC regions in the reported setting.Using only the extraction network performed much worse than the full pOSAL method, and the ROI size relaxes location deviations when OD boundaries are unclear.
  • Efficient architecture: MobileNetV2 is used as the backbone to reduce computation cost relative to Xception while retaining the segmentation framework’s efficiency focus.The cited comparison includes segmentation performance, parameter counts, and testing time cost.
  • Practical limitation: The method generalizes to unlabeled target-domain images, but training requires collecting such images and retraining for each new target domain.The authors identify domain generalization without many target images as future work because target images may be unavailable during training.

VI. CONCLUSION

The paper concludes that pOSAL combines efficient morphology-aware segmentation with patch-based output-space adaptation to address domain shift in OD and OC segmentation. Experiments on three public datasets are reported as demonstrating improvements and effectiveness, while extension to other medical imaging problems remains future work.

  • Conclusion: pOSAL segments optic disc and cup across fundus images using a lightweight network, morphology-aware loss, and unsupervised domain adaptation.Patch-based output-space adversarial learning captures local output-space statistics and guides similar source- and target-domain predictions.
  • Conclusion: Experiments on three public retinal fundus datasets demonstrate significant improvements and the effectiveness of the presented framework.
Loading 1902.07519v1…