No-reference quality index for color retinal images
Introduction
Retinal images are being increasingly used by ophthalmologists as well as in automatic systems for medical diagnosis and follow-up of retinal diseases such as diabetic retinopathy, glaucoma, and age-related macular degeneration. However, some retinal images can be unsuitable for reliable medical analysis and diagnosis due to their insufficient quality, most commonly due to reduced sharpness and/or uneven illumination. Poor quality retinal images can occur due to several factors including inadequate imaging conditions (e.g., insufficient illumination, poor focus) or patient related issues (e.g., pupil dilation, patient fixation, media opacity) [1], [2]. Several studies have shown that practical datasets can include a large number of poor quality retinal images that can be as high as 60% of the total images [3], [4].
Various shortcomings can result from employing bad quality retinal images in medical investigations. If a poor quality retinal image is not immediately identified, a recapture would be requested by the ophthalmologist costing himself and the patient both time and money, especially in cases when the imaging procedure and medical facility are in distant locations. Nevertheless, a worse scenario can occur if automatic screening systems are used for the medical diagnosis. Automatic retinal screening systems capture and process retinal images without any human intervention. Based on this analysis, the patient is advised to or against the need for further physical investigation depending on whether or not early disease symptoms were detected in the processed images. If poor quality retinal images are used within the automatic systems, misdiagnosis could occur leading to delayed treatment. As a result, further disease progressions can occur causing irreversible visual impairments that could lead to blindness.
No-reference retinal image quality assessment (RIQA) algorithms are being increasingly integrated as a preliminary preprocessing step in medical retinal image analysis in order to assure reliability of the performed diagnosis. RIQA algorithms automatically determine whether the captured images are suitable for reliable medical analysis in the absence of a gold standard reference image. Good quality retinal images are then stored and passed on for further processing, whereas poor quality images are either enhanced, if possible, to improve their quality or discarded altogether and an image recapture is performed. Generally, RIQA algorithms can be categorized into classification- and measure-based methods [5].
Classification-based RIQA algorithms rely on supervised learning methods to classify images into a specific quality class. Binary classification-based RIQA methods are very widely implemented in literature [3], [4], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15] where retinal images are considered to be either of excellent quality making them readily suitable for direct medical investigation or of severely poor quality rendering them unsuitable for medical diagnosis. However, captured retinal images may be of adequate quality which would require the application of proper image enhancement techniques (e.g., sharpening, luminosity improvement) before they are suitable for medical analysis. In this study, a practical dataset manually graded by human experts into good, adequate, and poor quality was found to have ∼45% of its images of adequate quality. In a previous work, Katuwal et al. [16] presented a multi-classification RIQA algorithm that categorized images into one of five different quality groups using several features related to the symmetry of the wavelet segmented blood vessels. The micro-averaged F-measure was found to be 0.6. The relatively low F-measure was attributed to the close similarity between the neighboring quality classes leading to increased misclassifications.
Measure-based RIQA algorithms compute a numeric quality index that is related to the quality of the retinal image. Numeric measures can help ophthalmologists associate a certainty degree to the diagnosis performed through inspection of a certain retinal image based on the value of the image's quality index [17]. Furthermore, numeric quality measures can facilitate evaluating the necessity and effectiveness of image enhancement methods performed to improve an image's quality making it more suitable for reliable medical diagnosis [17]. As a result, more adequate quality retinal images can be efficiently used, after proper enhancement, in medical analysis and diagnosis thus reducing the need for increased image recaptures. Numeric-based quality metrics have been commonly implemented in several fields including quality assessment of satellite images [18], stereoscopic images [19], [20], [21], underwater images [22], [23], generic images [24], [25], as well as for the evaluation of retinal image registration algorithms [26]. However, only few measure-based RIQA approaches exist in literature.
In the early work of Lee and Wang [17], a quality index was assigned to a test retinal image based on the similarity between its histogram and a template histogram created from a group of excellent quality retinal images. Although it was mentioned that their quality measure agreed with human perception, no correlation measure was given to quantify the degree of this agreement. In a more recent work, Bartling et al. [27] introduced a retinal image quality measure computed as the product of wavelet-based sharpness and illumination features. Retinal images were then classified into one of three quality categories based on intervals of the presented quality measure. The kappa value between the automatic evaluation and six human graders was found to be in the range of [0.52, 0.68]. Köhler et al. [5] presented a numeric RIQA index based on an adaptation of the general approach in which the retinal vessel tree was taken as guidance to determine a global quality score from local estimates in anisotropic patches [28]. Spearman's rank correlation coefficient between their proposed measure and each of the peak-signal-to-noise ratio (PSNR) and the structural similarity (SSIM) [29] full reference metrics was found to be 0.89 and 0.91, respectively. However, the PSNR has been reported to be inefficient in matching human judgement of image quality [30], [31] whereas the SSIM index was shown to be less competitive in assessing image quality related to its sharpness [32]. Measure-based retinal image quality techniques are thus still in their early stages which is indicated by the limited research and the relatively low statistical correlation, if any was given, between the existing measures and human experts.
In this work, a no-reference wavelet-based quality index is introduced to assess the quality of color retinal images based on their overall sharpness. Wavelet transform separates the sharpness and background information of an image in its detail and approximation subbands, respectively. Wavelet transform thus has the advantage of being consistent with the theory of human visual processing indicating that the eye has different optical paths for high and low frequencies [33]. Furthermore, wavelet multiresolution tends to bring out finer image details in the subsequent wavelet levels which were shown in previous work by Nirmala et al. [34] to be related to the different retinal structures. Recently, wavelet-based RIQA algorithms have been introduced to overcome the limitations in the more commonly utilized generic and segmentation based methods features by considering information related to the retinal structures and being computationally inexpensive while giving superior results [35]. Moreover, it was shown that transform-based retinal image quality features can be adapted to maintain reliable performance in practical scenarios in which the train and test datasets had significantly different image resolutions [36].
The introduced wavelet-based quality index is modified using a homogeneity parameter to account for reduced retinal image quality due to hidden structures in unevenly illuminated regions, thus increasing the reliability of the introduced measure. The homogeneity parameter was computed from the retinal saturation channel (Sretina) which was previously introduced by the authors to assess retinal image homogeneity [37].
In order to validate and test the proposed retinal image quality index, several analyses were performed. Two manually graded retinal image quality datasets having different resolutions, number of quality groups, and degree blurring of the bad quality images were included in these analyses. Initially, Spearman's rank correlation coefficient between the introduced quality index and the human graders was computed for the different datasets. Next, the retinal image quality index was input to a classifier to categorize the images into good, adequate, and bad quality retinal images. Finally, the presented quality index was employed to compare the performance of contrast limited adaptive histogram equalization (CLAHE), which is commonly implemented for retinal image contrast enhancement [38], [39], when it is applied to different color models. Several comparisons with other retinal image quality measures from literature are also presented showing superiority of the proposed index in reliability, performance, and computational time.
The rest of the paper is organized as follows: Section 2 summarizes the datasets used for devising and validating the proposed retinal image quality index. Section 3 presents the details of the retinal image quality measure and the theory behind its formation. Section 4 includes the results and discussion of the statistical tests and classification experiments performed to validate the introduced quality index. Moreover, several comparisons with other quality measures from literature are presented. Finally, Section 5 summarizes the conclusions of the presented work.
Section snippets
Materials
Several quality graded datasets were included in this study. The details of these datasets are as follows:
Dataset-1 (DS1) [40]: consists of 301 optic disc centered retinal images acquired with a Kowa nonmyd fundus camera and having a resolution of 1600 × 1212 pixels. Three human graders evaluated the quality of the images, then a majority vote determined their final quality class resulting in 236 good and 65 bad quality retinal images.
Dataset-2 (DS2): includes 190 optic disc centered retinal
Methods
Good quality retinal images should be sharp and evenly illuminated to facilitate the detection of retinal structures and possible disease lesions by both ophthalmologists and automatic systems. Wavelet transform has the ability to separate the sharpness information related to different retinal structures within the detail subbands whereas the low frequency information resides in the approximation subband. The introduced retinal image quality measure is hence computed from the wavelet
Results & discussion
Several experiments were performed to validate the efficiency of the introduced retinal image quality measure. Initially, the Spearman's rank correlation coefficient was computed to estimate the strength of association between the introduced quality index (Qrm) and the human graders. The proposed quality index was then used to classify retinal images into different quality groups and classification results are reported. Next, Qrm was calculated to compare the improvement induced in retinal
Conclusions
Numeric-based retinal image quality measures have several advantages over the more widely implemented classification-based approaches. Ophthalmologists can benefit from numeric quality measures in asserting a degree of confidence to the diagnosis performed based on examining a certain retinal image. Furthermore, automatic screening systems as well as imaging technicians can use the numeric quality index as a guidance to determine the degree of required image enhancement as well as to measure
Conflict of interest
None Declared.
References (49)
Automated retinal image quality assessment on the UK Biobank dataset for epidemiological studies
Comput. Biol. Med.
(2016)- et al.
Fully automated diabetic retinopathy screening using morphological component analysis
Comput. Med. Imaging Graph.
(2015) Quality assessment metric of stereo images considering cyclopean integration and visual saliency
Inf. Sci.
(2016)- et al.
CSV: image quality assessment based on color, structure, and visual system
Signal Process. Image Commun.
(2016) - et al.
Retinal blood vessel segmentation employing image processing and data mining techniques for computerized retinal image analysis
Biocybern. Biomed. Eng.
(2016) - et al.
Retinal image analysis: preprocessing and feature extraction
J. Phys. Conf. Ser.
(2011) - et al.
Progress towards automated diabetic ocular screening: a review of image analysis and intelligent systems for diabetic retinopathy
Med. Biol. Eng. Comput.
(2002) Automated image quality evaluation of retinal fundus photographs in diabetic retinopathy screening
Generic features for fundus image quality evaluation
Automatic no-reference quality assessment for retinal fundus images using vessel segmentation