Acquire Precise and Comparable Fundus Image Quality Score: FTHNet and FQS Dataset

TL;DR

A new FIQA dataset and model addressing the problems of current FIQA methods are established, and a FIQA Transformer-based Hypernetwork (FTHNet) is proposed to solve these tasks with regression results rather than classification results in conventional FIQA works.

Abstract

The retinal fundus images are utilized extensively in the diagnosis, and their quality can directly affect the diagnosis results. However, due to the insufficient dataset and algorithm application, current fundus image quality assessment (FIQA) methods are not powerful enough to meet ophthalmologists` demands. In this paper, we address the limitations of datasets and algorithms in FIQA. First, we establish a new FIQA dataset, Fundus Quality Score(FQS), which includes 2246 fundus images with two labels: a continuous Mean Opinion Score varying from 0 to 100 and a three-level quality label. Then, we propose a FIQA Transformer-based Hypernetwork (FTHNet) to solve these tasks with regression results rather than classification results in conventional FIQA works. The FTHNet is optimized for the FIQA tasks with extensive experiments. Results on our FQS dataset show that the FTHNet can give quality scores for fundus images with PLCC of 0.9423 and SRCC of 0.9488, significantly outperforming other methods with fewer parameters and less computation complexity.We successfully build a dataset and model addressing the problems of current FIQA methods. Furthermore, the model deployment experiments demonstrate its potential in automatic medical image quality control. All experiments are carried out with 10-fold cross-validation to ensure the significance of the results.

Figures (7)

• Figure 1: Exapmles of the Fundus Quality Score (FQS) dataset. In the FQS dataset, each fundus image has two labels: a three-level classification label (good, reject, and usable) and a continuous MOS varying from 0 to 100. Our FQS dataset covers the most common degradation types in clinical diagnoses, such as out-of-focus blur, haze, uneven illumination, and over-darkness.
• Figure 2: The statistic information of our FQS. (a) The MOS distribution histogram. Most of the MOSs are distributed between 60 and 80, which is consistent with actual clinical experience, and images of either extremely high or low quality are rare. (b) The standard deviation distribution histogram of MOSs. Half of the images have standard deviations under 4.34. Low SDs indicate that, though the opinion scores are given independently, the scoring criteria are consistent. (c) The three-level label distribution. The numbers of 'Good', 'Usable', and 'Reject' are 516, 793, and 937. There are more 'Reject' images to cover more degradation types.
• Figure 3: Architecture of our FTHNet. (a) The Transformer Backbone includes a patch embedding layer and four feature extraction stages. (b) The Distortion Perception Network is designed to extract distorted information. (c) The Parameter Hypernetwork comprises five parameter-generating layers. (d) The Target Network contains five linear layers to predict the fundus image quality scores. (e) The structure of the Basic Transformer Block. (f) The distortion perception block extracts the distortion information from the feature maps in different resolutions. (g) Each parameter-generating layer includes two branches to generate weight and bias parameters.
• Figure 4: Implementation of the FTHNet in the diagnosis system. The FTHNet works in the emphasized region.
• Figure 5: Failed cases of FTHNet on our FQS dataset. Most of these failed cases are from the "Reject" category. The FTHNet and FQS dataset will be refined according to these failed cases.