Source-linked AI summary
Evaluation of Retinal Image Quality Assessment Networks in Different Color-spaces
Huazhu Fu, Boyang Wang, Jianbing Shen, Shanshan Cui, Yanwu Xu, Jiang Liu, Ling Shao
TL;DR
RIQA lacks sufficiently large, multi-level, multi-modality evaluation data, while deep methods largely focus on RGB. The paper constructs EyeQ, analyzes color-spaces, and proposes MCF-Net with two-level fusion; experiments report superior RIQA performance and quality-dependent DR detection.
Problem
Existing RIQA methods focus on RGB and use small datasets with binary labels, while image quality varies across imaging conditions and affects diagnostic reliability.
Method
The paper re-annotates 28,792 EyePACS images into three quality grades and uses MCF-Net to fuse multiple color-space representations at feature and prediction levels.
Results
MCF-Net outperforms other methods, and automated DR detection performance is affected by image quality, including among images labeled Usable.
Takeaways & Limitations
EyeQ supports evaluation of RIQA and related applications, while image quality remains relevant to automated DR detection performance.
Abstract
from arXiv · showhide
Retinal image quality assessment (RIQA) is essential for controlling the quality of retinal imaging and guaranteeing the reliability of diagnoses by ophthalmologists or automated analysis systems. Existing RIQA methods focus on the RGB color-space and are developed based on small datasets with binary quality labels (i.e., `Accept' and `Reject'). In this paper, we first re-annotate an Eye-Quality (EyeQ) dataset with 28,792 retinal images from the EyePACS dataset, based on a three-level quality grading system (i.e., `Good', `Usable' and `Reject') for evaluating RIQA methods. Our RIQA dataset is characterized by its large-scale size, multi-level grading, and multi-modality. Then, we analyze the influences on RIQA of different color-spaces, and propose a simple yet efficient deep network, named Multiple Color-space Fusion Network (MCF-Net), which integrates the different color-space representations at both a feature-level and prediction-level to predict image quality grades. Experiments on our EyeQ dataset show that our MCF-Net obtains a state-of-the-art performance, outperforming the other deep learning methods. Furthermore, we also evaluate diabetic retinopathy (DR) detection methods on images of different quality, and demonstrate that the performances of automated diagnostic systems are highly dependent on image quality.
1 Introduction
RIQA is needed because retinal-image quality varies and can compromise diagnosis, while existing approaches and datasets have important limitations. The paper introduces a large, three-grade, multi-modality dataset and studies color-space fusion for RIQA.
- More than 25% of retinal images in a UK BioBank study were not high enough quality for accurate diagnosis.
- Quality degradation can prevent reliable diagnosis by ophthalmologists or automated analysis systems, making RIQA necessary.
- Existing RIQA methods include structure-based approaches dependent on segmentation and feature-based approaches using hand-crafted image or visual-system features.
- The paper re-annotates 28,792 EyePACS retinal images into Good, Usable, and Reject grades for evaluating RIQA and related applications.The dataset considers differences between ophthalmologists and automated systems.
- MCF-Net fuses representations from different color-spaces at feature and prediction levels, and experiments report that it outperforms other deep learning methods.
2 Eye-Quality Dataset
The EyeQ dataset addresses binary-label, single-modality, and small-size limitations in existing RIQA datasets. It uses three quality grades based on diagnosability across varied imaging conditions.
- Existing RIQA datasets commonly use binary Accept and Reject labels, despite images that are clinically gradable but potentially misleading for automated analysis.
- Existing datasets are often captured with one camera and are limited in size, leaving a lack of large-scale data for deep-learning quality grading.
- EyeQ re-annotates EyePACS images captured by different camera models and types under varied imaging conditions.
- The three grades consider blurring, uneven illumination, low contrast, and artifacts as common quality indicators.
- Good images have no low-quality factors and clearly visible retinopathy characteristics, whereas Usable images retain identifiable main structures and lesions despite slight quality problems.
- Reject images have serious quality issues preventing full and reliable diagnosis, including invisible disc or macula regions.
- Two experts graded EyePACS images, and ambiguous labels were discarded, yielding 28,792 retinal images.
3 Multiple Color-space Fusion Network
The paper studies color-space effects in deep RIQA networks and proposes MCF-Net to combine complementary representations. Its fusion operates at both feature and prediction levels with jointly weighted losses.
- Existing deep RIQA methods focus on RGB, while different color-spaces can represent diverse visual characteristics relevant to network performance.
- MCF-Net transfers RGB images into HSV and LAB, processes each color-space with a multi-scale CNN base network, and fuses their outputs.
- Feature-level fusion concatenates base-network feature maps and maps them to a fusion prediction through a fully connected layer.
- Prediction-level fusion concatenates the base-network predictions with the feature-level fusion prediction for the final quality output.
- MCF-Net combines base-network losses with feature- and prediction-level fusion losses, using weights 0.1, 0.1, and 0.6 to emphasize final prediction-level fusion.
4 Experiments
Experiments evaluate MCF-Net and individual color-space networks on EyeQ, then examine how image quality affects diabetic retinopathy detection. Results favor RGB/LAB representations, multi-color-space fusion, and higher-quality images.
- Experimental Settings: The EyeQ dataset uses 12,543 training images and 16,249 testing images to evaluate MCF-Net with ResNet18, ResNet50, and DenseNet121 base networks.Comparisons include RGB, HSV, LAB, and average prediction fusion.
- RIQA Results: Deep learning methods outperform the non-deep-learning baseline, while RGB and LAB networks perform better than HSV networks.The authors attribute these differences partly to RGB’s proximity to camera data and LAB’s lightness channel reflecting illumination.
- RIQA Results: Combining color-spaces outperforms individual color-spaces, including with simple average fusion.The result supports using complementary representations for retinal image quality assessment.
- RIQA Results: MCF-Net outperforms individual color-space models and average fusion, with DenseNet121-MCF achieving the best performance among the evaluated deep networks.The multi-level fusion block integrates information at feature and prediction levels.
- DR Detection: DR detection accuracy decreases as retinal image quality degrades across the evaluated ResNet18, ResNet50, and DenseNet121 models.The study evaluates these models on images grouped by quality using the EyeQ dataset.
- DR Detection: Poor-quality images remain challenging for automated diagnosis even when labeled Usable, although they provide diagnosable information to ophthalmologists.This distinguishes clinician usability from the demands of automated analysis.
5 Conclusion
The paper constructs the three-level EyeQ dataset and proposes MCF-Net for retinal image quality classification. Experiments show improved RIQA performance and demonstrate that image quality affects automated diabetic retinopathy detection.
- Dataset: EyeQ provides a large-scale retinal-image dataset with Good, Usable, and Reject quality grades.It is constructed from EyePACS and uses multi-level grading and multi-modality.
- Method: MCF-Net integrates different color-spaces for retinal image quality classification and outperforms other methods.The network is presented as a general approach for RIQA.
- Implications: Image quality affects the performance of automated diabetic retinopathy detection methods.The conclusion connects RIQA with automated disease diagnosis and related retinal-image applications.