An overview of the study approach and methodology is presented in Fig. 1.
A total of 1302 images were retrospectively collected to form 7 sub-datasets: DR-XJU, DR-ZJU, Glaucoma, Healthy, Local1, Local2 and UWF-mosaic. There are three types of images in these datasets: CFP images, portable camera images and UWF images. These images are from 904 patients, with ages ranging from 21 to 77 years. Written consent was signed by every participant before examinations to inform them that the images would be used for research purpose. Ethical approval for the study was obtained from the Ethics Committee of ZJU-2.
Specifically, DR-XJU, DR-ZJU, Glaucoma and Healthy subsets contained CFP images that were centerfield, including the optic disc and the macular area.
Images of DR-XJU were collected from patients with diabetic retinopathy at the Second Affiliated Hospital of Xi’an Jiaotong University (SAHXJU), captured with a Kowa non-mydriatic fundus camera (Kowa Company, Tokyo) with 45 degrees fields of view (FOV) and at 1924 by 1556 pixels.
Images of DR-ZJU, Glaucoma and Healthy were respectively collected from patients with diabetic retinopathy, glaucoma or no disease diagnosed at the Eye Center at the Second Affiliated Hospital of Zhejiang University School of Medicine (SAHZJU). The imaging device was a tabletop TRC-NW8 fundus camera (Top-Con Medical Systems, Tokyo) with 50 degrees FOV and a resolution of 1924 by 1556 pixels.
Local 1 and Local 2 contained portable camera images from healthy volunteers, and the imaging field included centerfield and other locations. These datasets were collected at the Eye Center at SAHZJU, captured with a DEC200 portable fundus camera (Med-imaging Integrated Solution Inc., Taiwan) with 60 degrees FOV and a resolution of 2560 by 1960 pixels. The difference between Local1 and Local2 was the imaging time period.
UWF-mosaic included UWF images from diabetic retinopathy patients. This dataset was also collected at the Eye Center at SAHZJU, and the capture device was an Optos ultra-wide field imaging system (Optos Plc Fife, Scotland) with 200 degrees FOV and a resolution of 1924 by 1556 pixels.
Detailed descriptions of the MSHF dataset is shown in Table 2. To show the diversity of the images in the MSHF dataset, we converted all images from the RGB color space to the Lab color space, and created a spatial scatter plot to show the distribution, as shown in Fig. 2. There are seven kinds of symbols in the figure, representing the seven subsets.
To facilitate the clinical application, the evaluation standard is a generic quality gradation scale that adhered to the generic-but-not-structural principle, as listed in Table 3. The overall quality represents the general impression of the images, and suggests whether or not the image is useable, while the illumination, clarity and contrast are parameters based on the characteristics of human visual system, and indicates the potential aspects to improve the image quality.
Images of the MSHF dataset were labelled by three ophthalmologists according to the principle. If the image was of good quality in a particular category, it was marked as ‘1’, and if not, as ‘0’. The ground truth was decided by the majority rule. Examples of high- and low-quality images of CFP, portable camera and UWF are shown in Fig. 3.
To make the MSHF dataset applicable for further AI model building, the dataset was manually divided into the training set (80%) and the test set (20%). The training set was used for learning and the test set was used for testing. There was no intersection between the 2 sets, and the variety of images was distributed equally. The 2 sets contained basically equal ratio of good- or poor-quality images. It is worth noting that we offered a possible way of split, and we did not mean to restrict its use. Future researchers can freely use this data set to achieve their research purposes.