Retinal image quality assessment based on image clarity and content

Enhancing Retinal Fundus Image Quality Assessment With Swin-Transformer-Based Learning Across Multiple Color-Spaces

Purpose

The assessment of retinal image (RI) quality holds significant importance in both clinical trials and large datasets, because suboptimal images can potentially conceal early signs of diseases, thereby resulting in inaccurate medical diagnoses. This study aims to develop an automatic method for Retinal Image Quality Assessment (RIQA) that incorporates visual explanations, aiming to comprehensively evaluate the quality of retinal fundus images (RIs).

Methods

We developed an automatic RIQA system, named Swin-MCSFNet, utilizing 28,792 RIs from the EyeQ dataset, as well as 2000 images from the EyePACS dataset and an additional 1,000 images from the OIA-ODIR dataset. After preprocessing, including cropping black regions, data augmentation, and normalization, a Swin-MCSFNet classifier based on the Swin-Transformer for multiple color-space fusion was proposed to grade the quality of RIs. The generalizability of Swin-MCSFNet was validated across multiple data centers. Additionally, for enhanced interpretability, a Score-CAM-generated heatmap was applied to provide visual explanations.

Results

Experimental results reveal that the proposed Swin-MCSFNet achieves promising performance, yielding a micro-receiver operating characteristic (ROC) of 0.93 and ROC scores of 0.96, 0.81, and 0.96 for the "Good," "Usable," and "Reject" categories, respectively. These scores underscore the accuracy of RIQA based on Swin-MCSF in distinguishing among the three categories. Furthermore, heatmaps generated across different RIQA classification scores and various color spaces suggest that regions in the retinal images from multiple color spaces contribute significantly to the decision-making process of the Swin-MCSFNet classifier.

Conclusions

Our study demonstrates that the proposed Swin-MCSFNet outperforms other methods in experiments conducted on multiple datasets, as evidenced by the superior performance metrics and insightful Score-CAM heatmaps.

Translational Relevance

This study constructs a new retinal image quality evaluation system, which will contribute to the subsequent research of retinal images.

Quality assessment of colour fundus and fluorescein angiography images using deep learning

BACKGROUND/AIMS

Image quality assessment (IQA) is crucial for both reading centres in clinical studies and routine practice, as only adequate quality allows clinicians to correctly identify diseases and treat patients accordingly. Here we aim to develop a neural network for automated real-time IQA in colour fundus (CF) and fluorescein angiography (FA) images.

METHODS

Training and evaluation of two neural networks were conducted using 2272 CF and 2492 FA images, with binary labels in four (contrast, focus, illumination, shadow and reflection) and three (contrast, focus, noise) modality specific categories plus an overall quality ranking. Performance was compared with a second human grader, evaluated on an external public dataset and in a clinical trial use-case.

RESULTS

The networks achieved a F1-score/area under the receiving operator characteristic/precision recall curve of 0.907/0.963/0.966 for CF and 0.822/0.918/0.889 for FA in overall quality prediction with similar results in most categories. A clear relation between model uncertainty and prediction error was observed. In the clinical trial use-case evaluation, the networks achieved an accuracy of 0.930 for CF and 0.895 for FA.

CONCLUSION

The presented method allows automated IQA in real time, demonstrating human-level performance for CF as well as FA. Such models can help to overcome the problem of human intergrader and intragrader variability by providing objective and reproducible IQA results. It has particular relevance for real-time feedback in multicentre clinical studies, when images are uploaded to central reading centre portals. Moreover, automated IQA as preprocessing step can support integrating automated approaches into clinical practice.

Multiresolution analysis for COVID-19 diagnosis from chest CT images: wavelet vs. contourlet transforms

Chest computer tomography (CT) provides a readily available and efficient tool for COVID-19 diagnosis. Wavelet and contourlet transforms have the advantages of being localized in both space and time. In addition, multiresolution analysis allows for the separation of relevant image information in the different subbands. In the present study, transform-based features were investigated for COVID-19 classification using chest CT images. Several textural and statistical features were computed from the approximation and detail subbands in order to fully capture disease symptoms in the chest CT images. Initially, multiresolution analysis was performed considering three different wavelet and contourlet levels to determine the transform and decomposition level most suitable for feature extraction. Analysis showed that contourlet features computed from the first decomposition level (L1) led to the most reliable COVID-19 classification results. The complete feature vector was computed in less than 25 ms for a single image having of resolution 256 × 256 pixels. Next, particle swarm optimization (PSO) was implemented to find the best set of L1-Contourlet features for enhanced performance. Accuracy, sensitivity, specificity, precision, and F-score of a 100% were achieved by the reduced feature set using the support vector machine (SVM) classifier. The presented contourlet-based COVID-19 detection method was also shown to outperform several state-of-the-art deep learning approaches from literature. The present study demonstrates the reliability of transform-based features for COVID-19 detection with the advantage of reduced computational complexity. Transform-based features are thus suitable for integration within real-time automatic screening systems used for the initial screening of COVID-19.

The Use of Datasets of Bad Quality Images to Define Fundus Image Quality

Screening programs for sight-threatening diseases rely on the grading of a large number of digital retinal images. As automatic image grading technology evolves, there emerges a need to provide a rigorous definition of image quality with reference to the grading task. In this work, on two subsets of the CORD database of clinically grad able and matching non-grad able digital retinal images, a feature set based on statistical and on task-specific morphological features has been identified. A machine learning technique has then been demonstrated to classify the images as per their clinical gradeability, offering a proxy for a rigorous definition of image quality.

Glaucoma Detection from Retinal Images Using Statistical and Textural Wavelet Features

Glaucoma is a silent progressive eye disease that is among the leading causes of irreversible blindness. Early detection and proper treatment of glaucoma can limit severe vision impairments associated with advanced stages of the disease. Periodic automatic screening can help in the early detection of glaucoma while reducing the workload on expert ophthalmologists. In this work, a wavelet-based glaucoma detection algorithm is proposed for real-time screening systems. A combination of wavelet-based statistical and textural features computed from the detected optic disc region is used to determine whether a retinal image is healthy or glaucomatous. Two public datasets having different resolutions were considered in the performance analysis of the proposed algorithm. An accuracy of 96.7% and area under receiver operating curve (AUC) of 94.7% were achieved for the high-resolution dataset. Analysis of the wavelet-based statistical and textural features using three different methods showed their relevance for glaucoma detection. Furthermore, the proposed algorithm is shown to be suitable for real-time applications as it requires less than 3 s for processing the high-resolution retinal images.

Robotic Ultrasound Scanning With Real-Time Image-Based Force Adjustment: Quick Response for Enabling Physical Distancing During the COVID-19 Pandemic

During an ultrasound (US) scan, the sonographer is in close contact with the patient, which puts them at risk of COVID-19 transmission. In this paper, we propose a robot-assisted system that automatically scans tissue, increasing sonographer/patient distance and decreasing contact duration between them. This method is developed as a quick response to the COVID-19 pandemic. It considers the preferences of the sonographers in terms of how US scanning is done and can be trained quickly for different applications. Our proposed system automatically scans the tissue using a dexterous robot arm that holds US probe. The system assesses the quality of the acquired US images in real-time. This US image feedback will be used to automatically adjust the US probe contact force based on the quality of the image frame. The quality assessment algorithm is based on three US image features: correlation, compression and noise characteristics. These US image features are input to the SVM classifier, and the robot arm will adjust the US scanning force based on the SVM output. The proposed system enables the sonographer to maintain a distance from the patient because the sonographer does not have to be holding the probe and pressing against the patient's body for any prolonged time. The SVM was trained using bovine and porcine biological tissue, the system was then tested experimentally on plastisol phantom tissue. The result of the experiments shows us that our proposed quality assessment algorithm successfully maintains US image quality and is fast enough for use in a robotic control loop.

An efficient framework for automated screening of Clinically Significant Macular Edema

The present study proposes a new approach to automated screening of Clinically Significant Macular Edema (CSME) and addresses two major challenges associated with such screenings, i.e., exudate segmentation and imbalanced datasets. The proposed approach replaces the conventional exudate segmentation based feature extraction by combining a pre-trained deep neural network with meta-heuristic feature selection. A feature space over-sampling technique is being used to overcome the effects of skewed datasets and the screening is accomplished by a k-NN based classifier. The role of each data-processing step (e.g., class balancing, feature selection) and the effects of limiting the region of interest to fovea on the classification performance are critically analyzed. Finally, the selection and implication of operating points on Receiver Operating Characteristic curve are discussed. The results of this study convincingly demonstrate that by following these fundamental practices of machine learning, a basic k-NN based classifier could effectively accomplish the CSME screening.

Ophthalmic image acquired by ophthalmologists and by allied health personnel as part of a telemedicine strategy: a comparative study of image quality

OBJECTIVES

This study evaluates the quality of ophthalmic images acquired by a nurse technician trained in teleophthalmology as compared with images acquired by an ophthalmologist, in order to provide a better understanding of the workforce necessary to operate remote care programs.

METHODS

A cross-sectional study was performed on 2044 images obtained from 118 participants of the TeleOftalmo project, in Brazil. Fundus and slit-lamp photography were performed on site by an ophthalmologist and by a nurse technician under the supervision of a remote ophthalmologist. Image quality was then evaluated by masked ophthalmologists. Proportion of suitable images in each group was compared.

RESULTS

The proportion of concordant classification regarding quality was 94.8%, with a corrected kappa agreement of 0.94. When analyzing each type of photo separately, there was no significant difference in the proportion of suitable images between on-site ophthalmologist and nurse technician with remote ophthalmologist assistance for the following: slit-lamp views of the anterior segment and anterior chamber periphery, and fundus photographs centered on the macula and on the optic disc (P = 0.825, P = 0.997, P = 0.194, and P = 0.449, respectively). For slit-lamp views of the lens, the proportion of suitable images was higher among those obtained by an ophthalmologist (99.6%) than by a technician (93.8%, P < 0.01).

CONCLUSIONS

Ophthalmic photographs acquired by a trained technician consistently achieved >90% adequacy for remote reading. Compared with ophthalmologist-acquired photos, the proportion of images deemed suitable achieved a high overall agreement. These findings provide favorable evidence of the adequacy of teleophthalmological imaging by nurse technicians.

A CNN-BASED RETINAL IMAGE QUALITY ASSESSMENT SYSTEM FOR TELEOPHTHALMOLOGY

Oculopathy is a widespread disease among people of all ages around the world. Teleophthalmology can facilitate the ophthalmological diagnosis for less developed countries that lack medical resources. In teleophthalmology, the assessment of retinal image quality is of great importance. In this paper, we propose a no-reference retinal image assessment system based on DenseNet, a convolutional neural network architecture. This system classified fundus images into good and bad quality or five categories: adequate, just noticeable blur, inappropriate illumination, incomplete optic disc, and opacity. The proposed system was evaluated on different datasets and compared to the applications based on other two networks: VGG-16 and GoogLenet. For binary classification, the good-and-bad binary classifier achieves an AUC of 1.000, and the degradation-specified classifiers that distinguish one specified degradation versus the rest achieve AUC values of 0.972, 0.990, 0.982, 0.982 for four categories, respectively. The multi-classification based on DenseNet achieves an overall accuracy of 0.927, which is significantly higher than 0.549 and 0.757 obtained using VGG-16 and GoogLeNet, respectively. The experimental results indicate that the proposed approach produces outstanding performance in retinal image quality assessment and is worth applying in ophthalmological telemedicine applications. In addition, the proposed approach is robust to the image noise. This study fills the gap of multi-classification in retinal image quality assessment.

Combination of Global Features for the Automatic Quality Assessment of Retinal Images

Diabetic retinopathy (DR) is one of the most common causes of visual loss in developed countries. Computer-aided diagnosis systems aimed at detecting DR can reduce the workload of ophthalmologists in screening programs. Nevertheless, a large number of retinal images cannot be analyzed by physicians and automatic methods due to poor quality. Automatic retinal image quality assessment (RIQA) is needed before image analysis. The purpose of this study was to combine novel generic quality features to develop a RIQA method. Several features were calculated from retinal images to achieve this goal. Features derived from the spatial and spectral entropy-based quality (SSEQ) and the natural images quality evaluator (NIQE) methods were extracted. They were combined with novel sharpness and luminosity measures based on the continuous wavelet transform (CWT) and the hue saturation value (HSV) color model, respectively. A subset of non-redundant features was selected using the fast correlation-based filter (FCBF) method. Subsequently, a multilayer perceptron (MLP) neural network was used to obtain the quality of images from the selected features. Classification results achieved 91.46% accuracy, 92.04% sensitivity, and 87.92% specificity. Results suggest that the proposed RIQA method could be applied in a more general computer-aided diagnosis system aimed at detecting a variety of retinal pathologies such as DR and age-related macular degeneration.

Retinal image quality assessment using deep learning

Poor-quality retinal images do not allow an accurate medical diagnosis, and it is inconvenient for a patient to return to a medical center to repeat the fundus photography exam. In this paper, a robust automatic system is proposed to assess the quality of retinal images at the moment of the acquisition, aiming at assisting health care professionals during a fundus photography exam. We propose a convolutional neural network (CNN) pretrained on non-medical images for extracting general image features. The weights of the CNN are further adjusted via a fine-tuning procedure, resulting in a performant classifier obtained only with a small quantity of labeled images. The CNN performance was evaluated on two publicly available databases (i.e., DRIMDB and ELSA-Brasil) using two different procedures: intra-database and inter-database cross-validation. The CNN achieved an area under the curve (AUC) of 99.98% on DRIMDB and an AUC of 98.56% on ELSA-Brasil in the inter-database experiment, where training and testing were not performed on the same database. These results show the robustness of the proposed model to various image acquisitions without requiring special adaptation, thus making it a good candidate for use in operational clinical scenarios.

No-reference quality index for color retinal images

Retinal image quality assessment (RIQA) is essential to assure that the images investigated by ophthalmologists or automatic systems are suitable for reliable medical diagnosis. Measure-based RIQA techniques have several advantages over the more commonly used binary classification-based RIQA methods. Numeric quality measures can aid ophthalmologists in associating a degree of confidence to the diagnosis performed through the investigation of a certain retinal image. Moreover, a numeric quality index can provide a mean for identifying the degree of enhancement required as well as to evaluate and compare the improvement achieved by enhancement techniques. In this work, a no-reference retinal image sharpness numeric quality index is introduced that is computed from the wavelet decomposition of the images. In order to account for the obscured retinal structures in unevenly illuminated image regions, the quality index is modified by a homogeneity parameter calculated from the previously introduced retinal image saturation channel. The proposed quality index was validated and tested on two datasets having different resolutions and quality grades. A strong (Spearman's coefficient > 0.8) and statistically highly significant (p-value < 0.001) correlation was found between the introduced quality index and the subjective human scores for the two different datasets. Moreover, multiclass classification using solely the devised retinal image quality index as a feature resulted in a micro average F-measure of 0.84 and 0.95 using the high and low resolution datasets, respectively. Several comparisons with other retinal image quality measures demonstrated superiority of the proposed quality index in both performance and speed.