[Paper Review] Retinal Optic Disc Segmentation using Conditional Generative Adversarial Network

Vivek Kumar Singh, Hatem A. Rashwan+8 more|arXiv (Cornell University)|Jun 11, 2018

Retinal Imaging and Analysis10 references19 citations

TL;DR

This paper proposes a conditional Generative Adversarial Network (cGAN) for retinal optic disc segmentation, using a U-Net-like encoder-decoder generator with skip connections and a discriminator to improve boundary accuracy. The method achieves state-of-the-art performance, attaining a Jaccard index of 0.96 on DRISHTI GS1 and 0.93 on RIM-ONE, with Dice scores of 0.97 and 0.98, respectively, while segmenting images in under a second on a GPU.

ABSTRACT

This paper proposed a retinal image segmentation method based on conditional Generative Adversarial Network (cGAN) to segment optic disc. The proposed model consists of two successive networks: generator and discriminator. The generator learns to map information from the observing input (i.e., retinal fundus color image), to the output (i.e., binary mask). Then, the discriminator learns as a loss function to train this mapping by comparing the ground-truth and the predicted output with observing the input image as a condition.Experiments were performed on two publicly available dataset; DRISHTI GS1 and RIM-ONE. The proposed model outperformed state-of-the-art-methods by achieving around 0.96% and 0.98% of Jaccard and Dice coefficients, respectively. Moreover, an image segmentation is performed in less than a second on recent GPU.

Motivation & Objective

• To develop a deep learning-based method for accurate and efficient segmentation of the optic disc in retinal fundus images.
• To address the challenge of precise optic disc boundary detection, crucial for diagnosing glaucoma and other retinal diseases.
• To leverage conditional GANs to improve segmentation quality by learning realistic texture and structural features from input fundus images.
• To outperform existing state-of-the-art methods in segmentation accuracy and robustness across diverse datasets.
• To enable real-time inference suitable for clinical deployment using modern GPU hardware.

Proposed method

• The proposed framework uses a conditional GAN with a generator network based on an encoder-decoder architecture, incorporating batch normalization and LeakyReLU activation functions.
• Skip connections are implemented between corresponding encoder and decoder layers to preserve spatial details and enhance feature learning.
• The generator maps input retinal fundus images to binary optic disc segmentation masks using transposed convolutions for up-sampling and a sigmoid activation in the final layer.
• The discriminator network evaluates the realism of the generated segmentation masks by comparing them with ground-truth masks, conditioned on the input image.
• The overall training objective combines binary cross-entropy loss with an adversarial loss to encourage the generator to produce masks indistinguishable from ground truth.
• The model is trained end-to-end on two public datasets: DRISHTI GS1 and RIM-ONE, using standard data augmentation and optimization techniques.

Experimental results

Research questions

• RQ1 Can a conditional GAN framework achieve superior optic disc segmentation performance compared to existing deep learning methods?
• RQ2 How does the inclusion of adversarial training and skip connections affect segmentation accuracy and boundary precision?
• RQ3 To what extent does the proposed cGAN generalize across diverse retinal image datasets with varying image quality and optic disc appearance?
• RQ4 Can the model achieve high accuracy while maintaining inference speed suitable for real-time clinical applications?
• RQ5 Does the adversarial component improve the model's ability to learn fine-grained structural details of the optic disc?

Key findings

• The proposed cGAN model achieved a Jaccard index of 0.96 and Dice coefficient of 0.97 on the DRISHTI GS1 dataset, outperforming all baseline methods.
• On the RIM-ONE dataset, the model achieved a Jaccard index of 0.93 and Dice coefficient of 0.98, demonstrating strong generalization across datasets.
• The model recorded an accuracy of 0.98, sensitivity of 0.98, and specificity of 0.99 on DRISHTI GS1, indicating high overall segmentation reliability.
• Inference time was less than one second per image on a modern GPU, confirming real-time feasibility.
• Qualitative results in Figure 4 show that the cGAN-generated masks closely match ground truth with sharp, accurate boundaries, surpassing FCN, SegNet, and U-Net.
• The method outperformed state-of-the-art approaches, including U-Net and methods by Shankaranarayana et al. and Zilly et al., in both quantitative metrics and visual quality.