Patch-based Output Space Adversarial Learning for Joint Optic Disc and Cup Segmentation (1902.07519v1)

Abstract: 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.

• The paper presents the pOSAL framework that uses patch-based adversarial learning to jointly segment optic disc and cup regions, effectively addressing domain shift.
• It integrates a novel morphology-aware loss and patch-based discriminator within a lightweight MobileNetV2 architecture to improve segmentation accuracy and efficiency.
• Evaluations on multiple retinal datasets demonstrate state-of-the-art performance, offering a scalable approach to enhance glaucoma diagnosis in clinical settings.

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

The paper "Patch-based Output Space Adversarial Learning for Joint Optic Disc and Cup Segmentation" by Shujun Wang et al. presents an innovative approach to optic disc (OD) and cup (OC) segmentation in retinal fundus images, leveraging deep learning techniques optimized for robust domain adaptation. This research primarily addresses the domain shift problem prevalent in medical imaging datasets, which leads to substantial degradation in segmentation performance when models are applied across datasets obtained through different imaging conditions or devices.

Core Contributions

The authors introduce the Patch-based Output Space Adversarial Learning (pOSAL) framework, which notably integrates several cutting-edge components:

1. Morphology-aware Segmentation Loss: The proposed segmentation network utilizes a novel morphology-aware loss function that incorporates both Dice coefficient loss and a smoothness loss. This design aims to facilitate accurate segmentation by respecting morphological constraints specific to the anatomical structures being imaged, i.e., ensuring the spatial correctness of OD and OC.
2. Patch-based Discriminator: Moving away from conventional domain adaptation methods that focus on feature space, the paper applies adversarial learning directly in the output space, utilizing a patch-based discriminator. This component plays a crucial role in aligning the output space distributions between source and target domains, enabling the network to mitigate domain shifts effectively.
3. Lightweight Architecture: The use of a MobileNetV2 backbone for the segmentation network reduces computational demand while maintaining accuracy. This optimization signifies a thoughtful balance between model complexity and performance, particularly beneficial for real-world applications requiring scalability and deployability in diverse clinical environments.

Methodological Insights

The implementation of the pOSAL framework effectively exploits the spatial and morphological consistencies inherent to the OD and OC across different datasets. This is complemented by unsupervised domain adaptation, ensuring that the model retains high performance even when confronted with new, unannotated datasets, thereby circumventing the typically laborious and costly annotation process.

Evaluations were conducted on three public retinal image datasets: Drishti-GS, RIM-ONE-r3, and REFUGE, where pOSAL achieved state-of-the-art segmentation results. Notably, in the MICCAI 2018 Retinal Fundus Glaucoma Challenge, the framework exhibited leading performance, reinforcing its practical competence in glaucoma screening.

Implications and Future Directions

This research provides significant contributions to the field of medical image analysis, particularly in addressing the robust domain adaptation challenge. The combination of adversarial learning with a morphology-aware loss function introduces a promising paradigm for medical applications where anatomical accuracy is paramount.

In future work, exploring domain generalization techniques that do not require prior access to target domain data could further expand the utility of this approach. Additionally, adaptations to accommodate other medical imaging domains could generalize this framework beyond retinal image segmentation.

In conclusion, this paper presents a significant advancement in the automated segmentation of optical components for glaucoma diagnosis. By aligning domain-independent features in the output space, the pOSAL framework paves the way for substantial improvements in the robustness and adaptability of deep learning models in medical contexts.