RSAP-Net: joint optic disc and cup segmentation with a residual spatial attention path module and MSRCR-PT pre-processing algorithm

Yun Jiang, Zeqi Ma, Chao Wu, Zequn Zhang, Wei YanBMC Bioinformatics 2022 · DOI

TL;DR

This paper introduces RSAP-Net, a novel deep learning architecture for simultaneously segmenting the optic disc (OD) and optic cup (OC) in fundus images. By incorporating a residual spatial attention path (RSAP) and a multi-scale Retinex color restoration with polar transformation (MSRCR-PT) preprocessing step, RSAP-Net achieves state-of-the-art performance in glaucoma screening by improving segmentation accuracy and boundary definition.

What Problem Does This Paper Address?

Glaucoma, a leading cause of irreversible blindness, often has no early symptoms. Early detection is crucial, and a key indicator is the assessment of the optic nerve head (ONH), specifically the optic disc (OD) and optic cup (OC). The cup-to-disc ratio (CDR) derived from segmenting these structures is a vital diagnostic metric. Manual segmentation is time-consuming, subjective, and not scalable for widespread screening. While deep learning methods, particularly Convolutional Neural Networks (CNNs), have shown promise, existing approaches often treat OD and OC segmentation as separate tasks, neglecting the inherent spatial relationship where the OC is contained within the OD. This oversight, coupled with challenges like blurred boundaries and varying image quality (illumination, color), limits segmentation accuracy.

Methods

The core of this work is the RSAP-Net, a U-shaped encoder-decoder architecture. The novelty lies in replacing the standard skip connections with a Residual Spatial Attention Path (RSAP). This RSAP module comprises residual convolutional blocks and a spatial attention mechanism. It aims to:

  • Mitigate feature discrepancies between the encoder and decoder, which arise from progressive downsampling and upsampling.
  • Aggregate global contextual information and focus on "where" features are most meaningful, thereby improving spatial domain understanding.
  • Enhance edge features of the OD and OC, crucial for accurate segmentation.

To further address challenges like blurred boundaries and the small size of the OC, a novel preprocessing method called MSRCR-PT is introduced. This method combines:

  • Multi-Scale Retinex Color Restoration (MSRCR): Enhances image contrast and color consistency, particularly beneficial for images with uneven illumination.
  • Polar Transformation (PT): Transforms the image into polar coordinates, which helps to normalize the cup-to-disk ratio and potentially improve segmentation by creating a more consistent representation.

The MSRCR-PT method is applied before feeding the image into the RSAP-Net.

Datasets

The primary dataset used for training and evaluation is the Drishti-GS1 standard dataset. This dataset contains 101 fundus images, with 50 for training and 51 for testing. It includes 70 images with glaucomatous lesions and 31 images of normal eyes. The images were acquired from patients aged 40-80 years and manually labeled by four experienced glaucoma specialists.

Key Insights and Ideas

  • Joint Segmentation is Key: The paper strongly advocates for treating OD and OC segmentation as a unified task, leveraging their inherent spatial relationship rather than as independent problems.
  • Attention for Feature Fusion: The RSAP module's success highlights the power of spatial attention in bridging the gap between encoder and decoder features, directly addressing a known limitation in U-Net-like architectures.
  • Preprocessing Matters Immensely: The MSRCR-PT method demonstrates that effective preprocessing can significantly boost the performance of even advanced segmentation models, particularly for challenging aspects like boundary definition and handling variations in image quality.
  • Polar Transformation's Role: While PT has been used before, its combination with MSRCR is shown to be particularly effective in normalizing the CDR and enhancing segmentation.

Key Results

RSAP-Net achieved state-of-the-art results on the Drishti-GS1 dataset:

  • OD Segmentation: F1 score of 0.9752, BLE of 6.33 pixels.
  • OC Segmentation: F1 score of 0.9012, BLE of 11.97 pixels.

These results are notably higher than many existing methods, including BGA-Net and JointRCNN. For OC segmentation, RSAP-Net's F1 score and BLE were 0.32% and 2.4 pixels better than the most advanced BGA-Net, respectively. Qualitatively, visual examples show RSAP-Net producing smoother and more accurate boundaries, especially for the OC, closely matching the ground truth. The model also demonstrates efficient inference times, with a segmentation map generated in just 1.32 seconds per image.

Limitations and Future Work

  • Task Specificity: The authors acknowledge that their current work focuses solely on OD and OC segmentation. They suggest investigating the generalizability of the MSRCR-PT preprocessing method and the RSAP module to other segmentation tasks.
  • Vessel Enhancement: While MSRCR has shown positive effects on fundus vessel enhancement in their experiments, this aspect was not the primary focus and warrants further exploration.
  • Other Segmentation Tasks: The paper explicitly states that the effectiveness of their method for tasks beyond OD/OC segmentation is yet to be demonstrated.

Summary updated

BibTeX
@article{jiang2022rsap,
  author = {Yun Jiang and Zeqi Ma and Chao Wu and Zequn Zhang and Wei Yan},
  title = {{RSAP-Net: joint optic disc and cup segmentation with a residual spatial attention path module and MSRCR-PT pre-processing algorithm}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-022-05058-2},
}
Figure