PALM: Open Fundus Photograph Dataset with Pathologic Myopia Recognition and Anatomical Structure Annotation
Extracted source text · arXiv 2305.07816
\documentclass[fleqn,10pt]{wlscirep}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
% \usepackage{lineno}
\usepackage{xcolor}
\usepackage{makecell}
% \linenumbers
\title{PALM: Open Fundus Photograph Dataset with Pathologic Myopia Recognition and Anatomical Structure Annotation}
\author[2,$\dag$]{Huihui Fang}
\author[1,$\dag$]{Fei Li}
\author[2]{Junde Wu}
\author[3]{Huazhu Fu}
\author[2]{Xu Sun}
\author[4]{Jos\'e Ignacio Orlando}
\author[5]{Hrvoje Bogunovi\'c}
\author[1,*]{Xiulan Zhang}
\author[2,*]{Yanwu Xu}
\affil[1]{ State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangdong Provincial Key Laboratory of Ophthalmology and Visual Science,Guangzhou, China}
\affil[2]{Intelligent Healthcare Unit, Baidu Inc., Beijing, China}
\affil[3]{Institute of High Performance Computing, Agency for Science, Technology and Research, Singapore}
\affil[4]{Yatiris Group, PLADEMA Institute, CONICET, UNICEN, Tandil, Argentina}
\affil[5]{Christian Doppler Lab for Artificial Intelligence in Retina, Department of Ophthalmology and Optometry, Medical University of Vienna, Vienna, Austria}
\affil[*]{corresponding authors: Yanwu Xu (xuyanwu@baidu.com); Xiulan Zhang (zhangxl2@mail.sysu.edu.cn)}
\affil[$\dag$]{these authors contributed equally to this work}
\begin{abstract}
Pathologic myopia (PM) is a common blinding retinal degeneration suffered by highly myopic population. Early screening of this condition can reduce the damage caused by the associated fundus lesions and therefore prevent vision loss. Automated diagnostic tools based on artificial intelligence methods can benefit this process by aiding clinicians to identify disease signs or to screen mass populations using color fundus photographs as inputs. This paper provides insights about PALM, our open fundus imaging dataset for pathological myopia recognition and anatomical structure annotation. Our databases comprises 1200 images with associated labels for the pathologic myopia category and manual annotations of the optic disc, the position of the fovea and delineations of lesions such as patchy retinal atrophy (including peripapillary atrophy) and retinal detachment. In addition, this paper elaborates on other details such as the labeling process used to construct the database, the quality and characteristics of the samples and provides other relevant usage notes.
\end{abstract}
\begin{document}
\flushbottom
\maketitle
% Click the title above to edit the author information and abstract
\thispagestyle{empty}
% \noindent Please note: Abbreviations should be introduced at the first mention in the main text – no abbreviations lists or tables should be included. Structure of the main text is provided below.
\section*{Background \& Summary}
%(700 words maximum)
Myopia has become a global burden of public health. In 2020, this condition affected nearly 30\% of the world population, and that number is expected to rise up to 50\% by 2050~\cite{sankaridurg2021imi}. Among myopic patients, about 10\% have high myopia~\cite{sankaridurg2021imi}, which is defined by a refractive error of at least -6.00D or an axial length of 26.5 mm or larger~\cite{percival1987redefinition}. As myopic refraction increases, there is an associated risk of pathological changes to the retina and choroid, i.e., high myopia will develop into pathological myopia (PM)~\cite{ohno2021imi}. PM is characterized by the formation of pathologic changes at the posterior pole and the optic disc and by myopic maculopathy~\cite{vingolo2019pathologic}. Among lesions usually observed in PM retinas, some of the more commonly seen are peripapillary atrophies, tessellations and macular hemorrhages (Figure.~\ref{fig:lesions}). These abnormalities can be observed in color fundus photography (CFP) (Figure~\ref{fig:lesions}), which is currently the most cost-effective imaging modality for this condition~\cite{orlando2020refuge}. As undetected PM might potentially result in irreversible visual impairment, it turns relevant to diagnose it at an early stage, to ensure regular patient follow-ups and treatments before further complications.
In view of the recent developments in artificial intelligence (AI) technology in the field of computer-aided disease diagnosis and treatment, multiple studies started to focus on applying this technology for automated analysis of CFPs~\cite{li2021applications,hagiwara2018computer,sengupta2020ophthalmic}. However, only a few studies aimed at PM in particular. We believe this is likely due to the fact that these data-driven models need to be trained using large curated and annotated datasets, which are currently scarce and not publicly available for this specific condition.
To facilitate future research in this topic, we provide PALM, an open database containing 1200 color fundus photographs related to PM~\cite{fu2019palm}. Unlike other disease datasets already available for the ophthalmic image analysis community such as SCES~\cite{baskaran2015prevalence}, ODIR~\cite{ODIR-5K} or AIROGS~\cite{de_vente_coen_2021_5745834}, ours includes not only CFPs and the disease labels but also optic disc segmentations, the location of the fovea and manual delineations of disease related lesions. These additional annotations can assist in building complementary AI models for disease classification and interpretation, which can aid clinicians to comprehensively analyze disease patterns and provide a more accurate diagnostic of PM.
%Moreover, the lesion annotations can also assist in the diagnosis and staging of pathologic myopia. In particular, as the clinicians prefer the availability of models that can assist in the interpretation of disease diagnostic results, by providing corresponding lesion detections, compared to a purely black-box automated disease diagnosis models.
PALM dataset has been released as part of the PAthologicaL Myopia challenge, which was held in conjunction with the International Symposium on Biomedical Imaging (ISBI) in 2019. To date, our dataset has already been used in more than 100 papers in the field of automated diagnosis of PM~\cite{biswas2022color,app11020591,cui2021pathological,rauf2021automatic,hemelings2021pathological} or fundus structure analysis~\cite{guo2020lesion,9193942,du2021deep} based on CFPs.
\section*{Methods}
\subsection*{Data Collection}
%The fundus image collection protocol was approved by the ethics board of Zhongshan Ophthalmic Center (ZOC), Sun Yat-sen University, China. Each retinal photography was obtained in a standardized dark room and with the patient sitting upright. Images were taken in a single field of view, i.e., the fundus was photographed with the midpoint of the optic disc (OD) and macula as the center, or in a dual field of view, i.e. with the OD and macula as the center of the image respectively (Figure.~\ref{fig:field_of_view}).
PALM contains retinal images retrospectively collected from a myopic examination cohort at the Zhongshan Ophthalmic Center (ZOC), Sun Yat-sen University, China. Each CFP was acquired in a single field of view, i.e., the fundus was photographed with the midpoint of the optic disc (OD) and macula as the center, or in a dual field of view, i.e. with the OD and macula as the center of the image, respectively (Figure.~\ref{fig:field_of_view}).
The protocol for retrieving the images was approved by the ethics board of ZOC. CFPs were included if: (1) they were acquired with a single field of view or with a dual field of view; (2) they show no image quality problems or minor exposure issues (minor smudges, out-of-focus or blurring that do not affect overall image interpretation were tolerated). % (3) corresponded to the left eye (\textit{oculus sinister}, OS).
Images were excluded if they showed any trace of treatment, severe exposure abnormalities, severe refractive interstitial opacities, large-scale contaminations or if information about its origin was missing.
%, which are selected from the previous myopic examination cohort in ZOC considering the following inclusion and exclusion criteria. Inclusion criteria: 1) CFPs taken with a single field of view or with a dual field of view; 2) No image quality problems or minor exposure problems, minor smudges that do not affect image interpretation, minor out-of-focus or blurring of images; 3) Oculus sinister (OS). Exclusion criteria: 1) Presence of treatment traces; 2) Severe exposure abnormality, severe refractive interstitial opacity, large-scale contamination, missing information and irrelevant images, etc.
The final dataset comprises 1200 fundus images from 720 subjects (Male: 48.1\%, mean age: 37.5 $\pm$ 15.91), captured with either a Zeiss Visucam 500 camera (1047 images, with a resolution of $2124 \times 2056$ pixels) or a Canon CR-2 camera (153 images, with a resolution of $1444 \times 1444$ pixels). The database is provided already split in a training, a validation and a test set (Table~\ref{tab:data_intro}), with images belonging to the same patient assigned to the same set.
\subsection*{Disease Diagnosis}
Disease labels indicating presence or absence of PM were assigned to each scan based on its associated clinical records, which considered in a comprehensive manner the medical history, refractive error, fundus imaging reports, etc. The guidelines of the International Myopia Institute~\cite{flitcroft2019imi} were followed, so that an individual was considered as PM if structural changes in the posterior segment of the eye caused by an excessive axial elongation associated with myopia were observed, including posterior staphyloma, myopic maculopathy, and high myopia-associated glaucoma-like optic neuropathy. These alterations were observed during clinical examination using multiple imaging modalities, including optical coherence tomography (OCT), fluorescein angiography (FA), and OCT angiography (OCTA). Notice that non-PM images might not necessarily correspond to healthy individuals, as shown in Figure.~\ref{fig:non-pm}.
%According to the International Myopia Institute (IMI)~\cite{flitcroft2019imi}, PM is defined as follows: \textit{excessive axial elongation associated with myopia that leads to structural changes in the posterior segment of the eye (including posterior staphyloma, myopic maculopathy, and high myopia-associated optic neuropathy) and that can lead to loss of best corrected visual acuity.} In addition, the subdivided lesions of PM are shown in Table~\ref{tab:PM_lesion}. To observe the presence of these lesions in the posterior polar region, macular region, and optic disc region, various imaging modalities in addition to CFP were used during clinical examination. Namely, optical coherence tomography (OCT), fluorescein angiography (FA), and OCT angiography (OCTA). The PM disease status labels of all samples in the PALM dataset were derived from clinical diagnosis, which considered in a comprehensive manner the medical history, refractive error, fundus imaging reports, etc. It is important to note that the non-PM images in the PALM dataset are not necessarily images of healthy individuals; they could be images of patients with other non-PM eye diseases (as shown in Figure.~\ref{fig:non-pm}).
% 想说这么设计是为了考察其他疾病对PM疾病诊断的影响，但数据集里没有其他疾病的明确标签，所以不好分析，先不写这一点了。
\subsection*{Manual Annotations}
Manual delineations of the optic disc and fundus lesions and the annotation of the fovea localization (Figure.~\ref{fig:annotation}) were performed by seven ophthalmologists with an average experience of 8 years in the field (ranging from 5 to 10 years) and one senior ophthalmologist, with more than 10 years of experience, all of them part of ZOC staff (Figure.~\ref{fig:data_collection_labelling}). All ophthalmologists annotated the structures by themselves without having access to any patient information or knowledge of disease prevalence in the data. Details regarding the annotation protocol followed for each specific target are provided in the sequel.
\subsubsection*{Optic disc annotation}
Experts used a free annotation tool with capabilities for image review, zoom, contrast enhancement, and circle and ellipse fitting, to manually draw elliptical structures approximately covering the optic disc. Pixels within the fitted area were then mapped to a binary pixel-wise segmentation mask. Annotations of the same image performed by the seven different graders were merged into a single one by majority voting. The senior ophthalmologist then performed a quality check of this resulting mask to account for any potential mistakes. When errors in the annotations were observed, the senior ophthalmologist analyzed each of the seven masks, removed those that were considered erroneous and repeated the majority voting process with the remaining ones.
\subsubsection*{Fovea localization}
A tool that allows to manually set the position of the crosshair on an image was used to approximate the location of the fovea. The final annotation was produced by averaging the seven coordinates provided by the ophthalmologists, which was further reviewed by the senior ophthalmologist. In case certain fovea coordinates in the set were considered erroneous, this expert removed them and repeated the averaging process with the remaining ones.
\subsubsection*{Lesion annotation}
Two types of lesions related to PM were annotated on each image: patchy retinal atrophy (including peripapillary atrophy) and retinal detachment. Experts used the same annotation tool as for the optic disc, although using a closed curve to outline the lesions. Unlike the SUSTech-SYSU dataset~\cite{lin2020sustech}, a free-form closed curve was used to allow a more accurate approximation of lesion borders. The same revision process used for the optic disc mask was followed for lesion masks as well.
\subsection*{Data Validation}
Data quality was automatically verified using the Multiple Color-space Fusion Network (MCF-Net) approach by Fu \textit{et al.}~\cite{fu2019evaluation}, which classifies color fundus images into quality grades good, usable and reject based on different color-space representation at feature and prediction levels. Tables~\ref{tab:fu-subclass} and~\ref{tab:fu-subsets} indicate the number of images grouped by quality according to each disease label and for each split, respectively.
%We used a deep learning-based method to assess the quality of the color fundus images in PALM dataset and to verify their usability. We used Multiple Color-space Fusion Network (MCF-Net) by Fu et al.~\cite{fu2019evaluation}, which integrates the different color-space representations at both a feature-level and prediction-level, to predict image quality grades: Good, Usable, and Reject. For different data subsets and different categories, the number of color fundus images corresponding to the three image quality grades is shown in Table~\ref{tab:fu-subclass} and Table~\ref{tab:fu-subsets}.
According to Table~\ref{tab:fu-subclass}, 93.8\% of the images in the non-PM category are classified either as good or usable, while this number reduces to 5.8\% in the PM category. This is related to the characteristics of the PM images. As shown in Figure~\ref{fig:lesions}, retinal detachment, peripapillary atrophy, retinal atrophy and other lesions are common in the fundus images of patients with PM, which interfere with the recognition of anatomical structures such as the optic disc and the macula. Fu et.al.~\cite{fu2019evaluation} model associated images with invisible disc or macula to the ‘Reject’ grade, which is common in other similar studies~\cite{raj2019fundus,shao2017automated,fleming2006automated,niemeijer2006image}. While this explains why the number of "reject" images is observed in the non-PM class, this renders erroneous in a real-world setting, in which images like these should be processed by the subsequent AI models to produce adequate diagnostics. The significance of the propose PALM dataset is therefore increased, as it provides a series of scans that reproduces clinical use-cases currently ignored or neglected by the existing literature.
%Therefore, the resulting quality grades of these PM images in PALM were also predicted as 'Reject'. In many studies about automated analysis and diagnosis based on color fundus images, images that do not identify fundus structures are typically excluded from the analysis. However, in real-world diagnosis and treatment settings, often it will be difficult to observe the entire fundus structure. Therefore, it is important to include such images for AI development in order to facilitate the translation of such models to clinical use. %these images should not be removed from the dataset, which are used for the AI model training; otherwise, the AI model will never be able to correctly analyze these images.
%As a consequence, the proposed PALM dataset contains both high quality and relatively low quality images, which can ensure a realistic dataset for developing AI solutions aimed at clinical use-cases.
In addition, it can be seen from Table~\ref{tab:fu-subsets} that the image quality distribution in training, validation and testing subsets of the proposed PALM dataset is relatively consistent. This ensures that the posterior evaluation of the model will not suffer from unstratified sampling and distribution biases.% which can assist in the evaluation of the performance of AI models stratified by image quality.
\section*{Data Records}
PALM is uploaded to FigShare as a compressed file and will be made publicly available by the time of publication. All personal information that could be used to identify the patients was removed before preparation. Data is provided already partitioned in folders \textit{Training set}, \textit{Validation set} and \textit{Test set}, with each subset containing folders for '\textit{Images}', '\textit{Disc Masks}', and '\textit{Lesion Masks}', and three Excel files (i.e. '\textit{Classification Labels.xlsx}', '\textit{Fovea Localization.xlsx}', and '\textit{Supplementary Information.xlsx}').
%The PALM dataset is uploaded to FigShare storage as a compressed file and will be made public when the manuscript is published. All personally identifiable information has been removed. The dataset has been divided into three parts: \textit{Training set}, \textit{Validation set} and \textit{Test set}. As shown in Figure.~\ref{fig:data_store}, each subset has three folders (i.e. '\textit{Images}', '\textit{Disc Masks}', and '\textit{Lesion Masks}'), and three Excel files (i.e. '\textit{Classification Labels.xlsx}', '\textit{Fovea Localization.xlsx}', and '\textit{Supplementary Information.xlsx}').
The '\textit{Images}' folder within each subset contains 400 color fundus images each, stored in JPEG format, with 8 bits per color channel. Similarly, the '\textit{Disc Masks}' has all binary optic disc masks associated to each fundus picture, as BMP files, also with 8 bits per color channel. On the other hand, the ’\textit{Lesion Masks}’ folder contains two subfolders, corresponding to each lesion type target, namely, '\textit{Atrophy}' and '\textit{Detachment}', with binary annotations for patchy retinal atrophy and retinal detachment, respectively. The file format of the lesion segmentation masks is consistent with those of the optic disc masks.
%For each subset, in the folder '\textit{Images}', the 400 color fundus images are stored in JPEG format, with 8 bits per color channel. In the '\textit{Disc Masks}' folder, binary optic disc masks corresponding to the fundus images with optic discs are stored in BMP format, with 8 bits per color channel. The ’\textit{Lesion Masks}’ folder contains two subfolders, corresponding to the lesion types, namely, '\textit{Atrophy}' and '\textit{Detachment}'. These two subfolders contain the corresponding binary lesion masks of the original color fundus images with patchy retinal atrophy or retinal detachment. The file format of the lesion segmentation masks is consistent with those of the optic disc masks.
The '\textit{Classification Labels.xlsx}' file contains the labels for PM classification, with 1 representing PM and 0 no-PM. The '\textit{Fovea Localization.xlsx}' provides the x- and y- coordinates of the fovea. Notice that a coordinate $(0, 0)$ is used when the fovea is not visible in the associated image. Additional information, i.e., the equipments used for image acquisition and the type of photo centering, are provided in the ’\textit{Supplementary Information.xlsx}' file, using in one column 1 to denote Zeiss and 2 to denote Canon devices, and, in a second column, 1, 2, and 3 to indicate optic disc centered, fovea centered, and center at the midpoint of optic disc and fovea, respectively.
\section*{Technical Validation}
% 第一部分（数据没到位，暂时不讲）
% 计算不同标注者和最后金标准的误差；
% 计算不同标注者之间的一致性
% different annotators had high consistency/inter-rater agreement, ensuring the high quality of the annotations of the proposed dataset.
Tables~\ref{tab:detachment} and~\ref{tab:atropy} provide the proportion of pixels corresponding to regions with retinal detachment and patchy retinal atrophy, respectively, differentiating by each disease category, acquisition protocol and subset. As expected, no retinal detachment lesions were found in images acquired with the optic disc at the center of the field of view or in images of patients with no PM (Table~\ref{tab:detachment}). Patchy retinal atrophies, on the other hand, are observed in both PM and non-PM categories (Table~\ref{tab:atropy}), although their size is much larger in PM subjects. Furthermore, these lesions are more frequently observed in images with visible optic disc, which is expected considering that these lesions appear at the vecinity of this anatomical structure.
%To demonstrate the characteristics of the samples in the PALM dataset, we provide the proportion of the lesion (retinal detachment and patchy retinal atrophy) region in the image of different types of samples under disease categories and photo centering categories, as shown in Table~\ref{tab:detachment} and Table~\ref{tab:atropy}. As can be seen from Table~\ref{tab:detachment}, no retinal detachment lesions were found in the color fundus images taken with the optic disc as the center of the field of view and in the color fundus images of Non-PM. This is expected, because the color fundus images centered on optic disc mainly provide fundus conditions in the optic disc region, whereas retinal detachment is usually observed in non-optic disc regions. In terms of disease dimension, retinal detachment is a common complication of PM, hence it was not observed in the non-PM category images in our dataset. Table~\ref{tab:atropy} clearly reflects the patchy retinal atrophy characteristic on fundus images of PM and non-PM, i.e., the atrophic lesions cover on average about 15\% region of the fundus images corresponding to PM, regardless of the photo centering applied. Less patchy retinal atrophy was found in the non-PM fundus images, but it may be present near the optic disc region due to the presence of the peripapillary atrophy in the high myopic samples of the non-PM samples. Therefore, in Table~\ref{tab:atropy}, more patchy retinal atrophy coverage areas was detected in the optic disc centered images.
In addition to discussing the characteristics of the lesions in the images, we also counted the properties of the fovea position in the fundus images with respect to photo centering used. Table~\ref{tab:fovea_loc} shows the mean $[\bar{x}, \bar{y}]$ of the normalized coordinates for the fovea localization among the fundus images with different photo centering in PALM dataset. $\bar{x} = \frac{1}{N} \sum_{i=1}^{N}\frac{x_i}{W_i}$, $\bar{y} = \frac{1}{N} \sum_{i=1}^{N}\frac{y_i}{H_i}$, where $[x_i, y_i]$ is the coordinate of the fovea in the $i$th image, and $H_i$ and $W_i$ are the height and width of the image. $N$ is the total number of the samples in the corresponding categories. From the table, we can see that in fundus image centered on the optic disc, the fovea appears on the right side of the images, as fundus pictures correspond in all cases to left eyes. In the images centered on the macula, the fovea appears in the center, and in the images centered on the midpoint of the optic disc and the macula, the fovea appears to the right of the image center. Thus, the fovea position characteristics are consistent with our expectations.
\section*{Usage Notes}
Our dataset will be accessible through a FigShare link by the time of publication. %Our dataset can be downloaded from the link in this manuscript.
PALM images can be used to perform studies on automated PM classification, optic disc segmentation, fovea localization, and atrophic lesion retinal detachment segmentation. In the aforementioned PALM Challenge, these tasks were set up as sub-challenges in which different participating teams proposed their own methods to automate them. The evaluation of their corresponding approaches in the validation and test sets for each of the sub-challenges are accessible in \url{https://palm.grand-challenge.org/SemifinalLeaderboard/} and \url{https://palm.grand-challenge.org/Test/}.
For the studies on classification, segmentation and localization, we designed a series of baseline models~\cite{fang2022dataset}, which we trained and evaluated using PALM data. For optic disc and lesion segmentation, we used a standard U-shaped network~\cite{ronneberger2015u} with residual blocks, while for PM classication and fovea localization we utilized ResNet50~\cite{he2016deep} architectures. The corresponding code has been released as open source (See Code availability section). %In this section, we report the performance of the baseline models trained and evaluated with the PALM dataset on each task.
Table~\ref{tab:baseline_results} shows the evaluation results of the baseline models on each task. The chosen evaluation metrics are commonly used for classification, segmentation and localization tasks, and are equal to those used in other related ophthalmic image analysis challenges~\cite{orlando2020refuge, fang2022adam}. The baseline classification model achieved good results on both validation and test sets, as observed in Table~\ref{tab:baseline_results}. This is mainly because the fundus images of PM category in our dataset correspond to severe PM, so the associated fundus lesion features were particularly evident. For optic disc segmentation, results are slightly lower than those observed in cases with other conditions such as glaucoma~\cite{orlando2020refuge}. This is likely due to the PM cases showing significant abnormalities within the optic disc region. Furthermore, results for lesion segmentation achieved by the baseline model are poor. This might be a consequence of both the irregular shape of patchy retinal atrophy and retinal detachment lesions (which vary significantly from one case to another) and due to the scarcity of sufficient training samples. In the future, we will aim at collecting more PM images of various processes, and supplement the dataset with additional samples of patchy retinal atrophy and retinal detachment lesions to reduce the impact of this observation.
Tables~\ref{tab:categories_PM_class} to \ref{tab:categories_detach} show the performance of the baseline models stratified by category. As stated above, the baseline model was able to segment the optic disc region more accurately in non-PM samples than in PM cases (Table~\ref{tab:categories_OD_seg}). This is because PM lesions interfere with the overall appearance of the optic disc features, thus leading to poor segmentation prediction results. Therefore, new models should incorporate mechanisms to deal with these characteristics. A similar conclusion can be drawn for the fovea localization task (Table~\ref{tab:categories_fovea}), in which much larger errors are observed in PM cases. This further emphasize the importance of PALM, which complements other existing databases with much more challenging scenarios to further improve existing models. Results for patchy retinal atrophy segmentation (Table~\ref{tab:categories_atrophy}) are much better in PM cases than in non-PM, which can be due to the much more subtle appearance of these lesions in non-PM. Finally, notice that performance in the validation and test sets are similar (Tables~\ref{tab:categories_PM_class}-\ref{tab:categories_detach}), which indicates that our data partition is well-balanced, and suitable for researchers to directly use to tune and test the models.
%dataset can complement other datasets for building AI systems for optic disc segmentation or fovea localization, because the lesion areas in the PM-category fundus images of the PALM dataset present a great challenge for automated analysis of the ocular structures, and can serve as a source of hard samples for the training process.
%In addition, as can be seen from Table~\ref{tab:categories_fovea}, fovea localization results were slightly worse for CANON and Optic disc centered categories. This is due to the fact that most CANON samples were also optic disc centered samples in our dataset, and the maculas were incomplete in the images centered on the optic disc, creating difficulties for locating the fovea. From Table~\ref{tab:categories_atrophy}, we can see that the segmentation accuracy of atrophic lesions in PM samples was better than that in non-PM samples, which is due to the more obvious lesion area in PM samples. In addition, it can be found that the prediction accuracy on the validation set and the test set in Tables~\ref{tab:categories_PM_class}-\ref{tab:categories_detach} were similar, which indicates that our data partition is well-balanced, and suitable for researchers to directly use to tune and test the models.
In summary, PALM is the first dataset for assisting AI researchers in training AI models for automated PM analysis. Disease labels are complemented by a series of manual annotations of lesions and anatomical structures that can allow studies focusing on exploiting complementary features to enhance results. Furthermore, PALM can be used in combination with other existing fundus image datasets such as REFUGE~\cite{orlando2020refuge} and ADAM~\cite{fang2022adam} to produce much more robust models for optic disc segmentation, fovea localization and even quality assessment.
%Following the above discussion, the PALM dataset allows us to evaluate the predictive performance of AI models across multiple important and clinically-relevant aspects. At the same time, it also enlightens us that the training of AI models can be carried out separately for different kinds of data. In addition, the PALM dataset can be used in combination with other similar open fundus image datasets, such as REFUGE~\cite{orlando2020refuge}, and ADAM~\cite{fang2022adam}.
\section*{Code availability}
The source code for the image quality assessment by \textit{Fu et.al.} can be accessed at \url{https://github.com/hzfu/EyeQ}. The source code for the baseline model training and testing is available at \url{https://github.com/tianyizheming/ichallenge_baseline}.
\bibliography{sample}
\section*{Author contributions statement}
Acquisition of data: X.Z., and F.L.\\
Analysis and interpretation of data: H.F. (Huihui Fang), J.W., and X.S.\\
PALM Challenge organization: Y.X., X.Z., H.F. (Huazhu Fu), F.L., X.S., J.I.O, H.B.\\
Drafting the work or revising it critically: H.F. (Huihui Fang), F.L., J.I.O, H.B., H.F. (Huazhu Fu), J.W., and Y.X.
\section*{Competing interests}
The authors declare no competing interests.
\section*{Figures \& Tables}
% 一个展示最终标签的图（GT渲染图+GT真实状态图，看情况设计）
% Technical validation部分：一个展示不同医生标注结果的图（三个标注任务按照标注人以不同颜色渲染）（可以病理数据和非病理数据各选一张做展示，2行）（数据未到位，暂不写这部分）
% Technical validation部分：按照上述根据数据属性的讨论，给不同类别下的分割GT效果
% note usage部分：展示baseline在不同数据属性样本上的分割or定位效果
\begin{figure}[h!]
\centering
\includegraphics[width=0.75\linewidth]{figures/lesions_1006.pdf}
\caption{Examples of retinal lesions commonly observed in PM cases: (a) Peripapillary atrophies, which occur at the proximity of the optic disc; (b) Tessellated retina, with an observable large choroidal vessels at the posterior fundus pole; (c) Macular hemorrhages, mostly along the crack itself and near from the center of the fovea or in its immediate vicinity; (d) Retinal atrophy, pigment clumping in and around the lesion due to migration of the degenerated retinal pigment epithelium cells into the inner retinal layers; (e) Retinal detachment, an emergency situation in which the retina is pulled away from its normal position; (f) Vitreous opacity, in which the vitreous shrinks and forms strands that cast shadows on the retina. All images corresponds to training samples from PALM.}
\label{fig:lesions}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.75\linewidth]{figures/field_of_view.pdf}
\caption{Color fundus images from PALM training set centered at (a) the midpoint between optic disc and fovea, (b) the optic disc, and (c) the fovea).}
\label{fig:field_of_view}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/non-pm-show.pdf}
\caption{Examples of color fundus images from PALM corresponding to the non-PM category. Notice that this subset contains not only healthy individuals (a) but also subjects with other conditions such as high myopia (b), epiretinal membrane (c) and diabetic retinopathy (d), among others.}
%\caption{Color fundus images of non-PM label}
\label{fig:non-pm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/labels_show.pdf}
\caption{Example of the annotation interface used by the experts in (a) a no-PM sample, and (b) a PM sample. (a1) and (b1): original input images, (a2) and (b2): manual annotations.}
\label{fig:annotation}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\linewidth]{figures/labeling_1006.pdf}
\caption{Manual annotation process. Manual delineations were performed by seven different experts and reviewed subsequently by one senior expert.
}
\label{fig:data_collection_labelling}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.35\linewidth]{figures/data_store.pdf}
\caption{Folder organization of our PALM dataset.}
\label{fig:data_store}
\end{figure}
\begin{table}[h!]
\centering
\caption{Summary of the main characteristics of each subset of the PALM dataset, stratified by disease, structure, lesion, image acquisition type, and acquisition device.}
\begin{tabular}{ccccccccc}
\hline
Set & Num. & PM/Non-PM & With/o OD & With/o Fovea & \makecell[c]{With/o \\Detachment} & \makecell[c]{With/o \\ Atrophy} & \makecell[c]{Photo centering \\(OD/ fovea/ \\ midpoint of \\ OD and fovea)} & \makecell[c]{Device \\(Zeiss/Canon)} \\
\hline
Training & 400 & 213/187 & 381/19 & 397/3 & 12/388 & 311/89 & 42/258/100 & 350/50 \\
Validation & 400 & 211/189 & 379/21 & 397/3 & 6/394 & 271/129 & 43/258/99 & 344/56 \\
Testing & 400 & 213/187 & 384/16 & 398/2 & 6/394 & 288/112 & 38/284/78 & 353/47 \\
\hline
Total & 1200 & 637/563 & 1144/56 & 1192/8 & 24/1176 & 870/330 & 123/800/277 & 1047/153 \\
\hline
\end{tabular}%
\label{tab:data_intro}%
\end{table}%
% \begin{table}[h!]
% \centering
% \caption{Subdivisions of PM by International Myopia Institute (IMI)~\cite{flitcroft2019imi}}
% \begin{tabular}{cc}
% \hline
% Term & Definition \\
% \hline
% % \makecell[c]{Pathologic myopia} & \makecell[c]{Excessive axial elongation associated with myopia that leads to structural\\ changes in the posterior segment of the eye (including posterior staphyloma,\\ myopic maculopathy, and high myopia-associated optic neuropathy) and \\that can lead to loss of bestcorrected visual acuity.} \\
% \makecell[c]{Myopic macular degeneration} & \makecell[c]{A vision-threatening condition occurring in people with myopia, usually \\high myopia that comprises diffuse or patchy macular atrophy with or without \\lacquer cracks, macular Bruch’s membrane defects, CNV and Fuchs spot.}\\
% \makecell[c]{Myopic traction maculopathy } & \makecell[c]{A combination of macular retinoschisis, lamellar macula hole and/or \\foveal retinal detachment (FRD) in eyes with high myopic attributable to\\ traction forces arising from adherent vitreous cortex, epiretinal membrane, \\internal limiting membrane, retinal vessels, and posterior staphyloma.} \\
% \makecell[c]{Myopia-associated glaucoma-like\\ optic neuropathy} & \makecell[c]{Optic neuropathy characterized by a loss of neuroretinal rim and \\enlargement of the optic cup, occurring in eyes with high myopia \\eyes with a secondary macrodisc or peripapillary delta zone at a normal IOP. }\\
% \hline
% \end{tabular}%
% \label{tab:PM_lesion}%
% \end{table}%
\begin{table}[h!]
\begin{minipage}[c]{0.5\textwidth}
\centering
\captionof{table}{The image quality assessment results in PM and Non-PM samples of the PALM dataset according to the fundus image quality assessment method proposed by Fu et al.~\cite{fu2019evaluation}}
\label{tab:fu-subclass}
\begin{tabular}{cccc}
\hline
Quality & Good & Usable & Reject \\
\hline
PM & 37 & 123 & 477 \\
Non-PM & 238 & 290 & 35 \\
\hline
\end{tabular}
\end{minipage}
\begin{minipage}[c]{0.5\textwidth}
\centering
\captionof{table}{The image quality assessment results in different subsets of the PALM dataset according to the fundus image quality assessment method proposed by Fu et al.~\cite{fu2019evaluation}}
\label{tab:fu-subsets}
\begin{tabular}{cccc}
\hline
Quality & Good & Usable & Reject \\
\hline
Training set & 84 & 143 & 173 \\
Validation set & 87 & 143 & 170 \\
Testing set & 104 & 127 & 169 \\
\hline
\end{tabular}
\end{minipage}
\end{table}
% 数据集中PM样本视盘区域占比变小不明显
% \begin{table}[htbp]
% \centering
% \caption{The proportion of pixels of optic disc in different categories of fundus images in PALM dataset (\%)}
% \begin{tabular}{ccccccc}
% \hline
% & \multicolumn{2}{c}{Training} & \multicolumn{2}{c}{Validation} & \multicolumn{2}{c}{Testing} \\
% \cline{2-7}
% & \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} & \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} & \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} \\
% \hline
% OD\_CEN & 3.305 & 2.469 & 3.524 & 2.985 & 3.06 & 2.641 \\
% FOVEA\_CEN & 1.492 & 1.6 & 1.45 & 1.61 & 1.403 & 1.568 \\
% MIDPOINT\_CEN & 1.666 & 1.649 & 1.635 & 1.894 & 1.602 & 1.523 \\
% \hline
% \end{tabular}%
% \label{tab:od_\%}%
% \end{table}%
\begin{table}[htbp]
\centering
\caption{Proportion of the detachment mask pixels in different categories of fundus images in the PALM dataset.}%The proportion of retinal detachment pixels by different categories of fundus images in PALM dataset.}
\begin{tabular}{ccccccc}
\hline
& \multicolumn{2}{c}{Training} & \multicolumn{2}{c}{Validation} & \multicolumn{2}{c}{Testing} \\
\cline{2-7}
& \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} & \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} & \multicolumn{1}{c}{PM} & \multicolumn{1}{c}{Non-PM} \\
\hline
Optic disc centered & 0\% & 0\% & 0\% & 0\% & 0\% & 0\% \\
Fovea centered & 3.6\% & 0\% & 1.2\% & 0\% & 0.8\% & 0\% \\
\makecell[c]{Midpoint of the optic disc\\ and fovea centered} & 2.4\% & 0\% & 2.3\% & 0\% & 3.6\% & 0\% \\
\hline
\end{tabular}%
\label{tab:detachment}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Proportion of the atrophy mask pixels in different categories of fundus images in the PALM dataset.}%The proportion of atrophic pixels by different categories of fundus images in PALM dataset.}
\begin{tabular}{ccccccc}
\hline
& \multicolumn{2}{c}{Training} & \multicolumn{2}{c}{Validation} & \multicolumn{2}{c}{Testing} \\
\cline{2-7}
& PM & Non-PM & PM & Non-PM & PM & Non-PM \\
\hline
Optic disc centered & 15.1\% & 0.6\% & 16.8\% & 1.1\% & 16.8\% & 1.5\% \\
Fovea centered & 9.8\% & 0.3\% & 14.3\% & 0.3\% & 13.3\% & 0.3\% \\
\makecell[c]{Midpoint of the optic disc\\ and fovea centered} & 14.2\% & 0.3\% & 15.7\% & 0.3\% & 20.3\% & 0.5\% \\
\hline
\end{tabular}%
\label{tab:atropy}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{The mean $[\bar{x}, \bar{y}]$ of the normalized coordinates for the fovea localization among the fundus images with different photo centering in PALM dataset.}
\begin{tabular}{cccc}
\hline
& Training & Validation & Testing \\
\hline
Optic disc centered & [0.856, 0.495] & [0.846, 0.506] & [0.851, 0.515] \\
Fovea centered & [0.522, 0.513] & [0.524, 0.516] & [0.5224, 0.514] \\
\makecell[c]{Midpoint of the optic disc \\ and fovea centered} & [0.614, 0.503] & [0.607, 0.503] & [0.607, 0.506] \\
\hline
\end{tabular}%
\label{tab:fovea_loc}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Performance of baseline classification, segmentation and localization models trained using PALM training set, as evaluated in PALM's validation and test sets. OD: Optic Disc. AUC: Area Under the receiver-operating Curve. F1: F1-score. Acc: Accuracy. Sen: Sensitivity. Spe: specificity. ED: Euclidean Distance.}
\begin{tabular}{cccccccccc}
\hline
& \multicolumn{5}{c}{Classification} & \makecell[c]{OD \\ Segmentation} & \makecell[c]{Fovea \\ Localization} & \makecell[c]{Atrophy \\ Segmentation} & \makecell[c]{Detachment \\Segmentation} \\
\cline{2-10}
& AUC & F1 & Acc & Sen & Spe & DICE & ED (pixels) & DICE & DICE\\
\hline
Validation & 0.999 & 0.981 & 0.980 & 0.976 & 0.984 & 0.876$\pm$0.148 & 90.928$\pm$134.942 & 0.675$\pm$0.266 & 0.412$\pm$0.284 \\
Test & 0.994 & 0.969 & 0.968 & 0.962 & 0.973 & 0.868$\pm$0.142 & 83.998$\pm$103.3.4 & 0.674$\pm$0.253 & 0.474$\pm$0.272 \\
\hline
\end{tabular}%
\label{tab:baseline_results}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Automated PM classification performance of our baseline trained using PALM and evaluated in the validation and test sets, stratified by disease, acquisition device and image acquisition type, in terms of accuracy (Acc). Acc for PM and Non-PM categories are calculated by $\frac{TP}{N_{PM}}$, and $\frac{TN}{N_{Non-PM}}$, respectively. $N_{PM}$ and $N_{Non-PM}$ are the total numbers of samples in the corresponding categories.}
\begin{tabular}{cccccccc}
\hline
Acc & PM & Non-PM & ZEISS & CANON & Optic disc centered & Fovea centered & \makecell[c]{Midpoint of the optic disc \\ and fovea centered} \\
\hline
Validation & 0.976 & 0.984 & 0.98 & 0.982 & 1 & 0.977 & 0.98 \\
Testing & 0.962 & 0.973 & 0.972 & 0.936 & 0.895 & 0.979 & 0.962 \\
\hline
\end{tabular}%
\label{tab:categories_PM_class}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Automated optic disc segmentation performance of our baseline trained using PALM and evaluated in the validation and test sets, stratified by disease, acquisition device and image acquisition type, in terms of DICE.}
\begin{tabular}{cccccccc}
\hline
& PM & Non-PM & ZEISS & CANON & \makecell[c]{Optic disc \\ centered} & \makecell[c]{Fovea\\ centered} & \makecell[c]{Midpoint of the\\ optic disc and\\ fovea centered} \\
\hline
Validation & 0.831$\pm$0.188 & 0.922$\pm$0.064 & 0.885$\pm$0.137 & 0.802$\pm$0.205 & 0.802$\pm$0.200 & 0.894$\pm$0.122 & 0.865$\pm$0.168 \\
Testing & 0.820$\pm$0.181 & 0.918$\pm$0.046 & 0.869$\pm$0.145 & 0.861$\pm$0.115 & 0.865$\pm$0.115 & 0.882$\pm$0.123 & 0.821$\pm$0.197 \\
\hline
\end{tabular}%
\label{tab:categories_OD_seg}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Automated fovea localization performance of our baseline trained using PALM and evaluated in the validation and test sets, stratified by disease, acquisition device and image acquisition type, in terms of Euclidean Distance (ED).}
\begin{tabular}{cccccccc}
\hline
ED (pixels) & PM & Non-PM & ZEISS & CANON & \makecell[c]{Optic disc \\ centered} & \makecell[c]{Fovea\\ centered} & \makecell[c]{Midpoint of the\\ optic disc and\\ fovea centered} \\
\hline
Validation & 108.1$\pm$178.3 & 71.8$\pm$48.5 & 80.8$\pm$106.7 & 153.2$\pm$235.8 & 188.3$\pm$339.4 & 71.1$\pm$74.5 & 100.2$\pm$62.0 \\
Testing & 103.4$\pm$132.8 & 62.0$\pm$42.7 & 77.3$\pm$92.6 & 134.2$\pm$153.6 & 119.6$\pm$62.1 & 76.0$\pm$114.6 & 95.8$\pm$63.5 \\
\hline
\end{tabular}%
\label{tab:categories_fovea}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Automated patchy retinal atrophy segmentation performance of our baseline trained using PALM and evaluated in the validation and test sets, stratified by disease, acquisition device and image acquisition type, in terms of DICE.}
\begin{tabular}{cccccccc}
\hline
DICE & PM & Non-PM & ZEISS & CANON & \makecell[c]{Optic disc \\ centered} & \makecell[c]{Fovea\\ centered} & \makecell[c]{Midpoint of the\\ optic disc and\\ fovea centered} \\
\hline
Validation & 0.778$\pm$0.182 & 0.358$\pm$0.232 & 0.655$\pm$0.276 & 0.751$\pm$0.209 & 0.796$\pm$0.149 & 0.619$\pm$0.287 & 0.717$\pm$0.243 \\
Testing & 0.770$\pm$0.171 & 0.407$\pm$0.254 & 0.656$\pm$0.263 & 0.770$\pm$0.163 & 0.772$\pm$0.159 & 0.613$\pm$0.269 & 0.779$\pm$0.193 \\
\hline
\end{tabular}%
\label{tab:categories_atrophy}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{Automated retinal detachment segmentation performance of our baseline trained using PALM and evaluated in the validation and test sets, stratified by disease, acquisition device and image acquisition type, in terms of DICE.}
\begin{tabular}{cccccccc}
\hline
DICE & PM & Non-PM & ZEISS & CANON & \makecell[c]{Optic disc \\ centered} & \makecell[c]{Fovea\\ centered} & \makecell[c]{Midpoint of the\\ optic disc and\\ fovea centered} \\
\hline
Validation & 0.412$\pm$0.284 & - & 0.412$\pm$0.284 & - & - & 0.549$\pm$0.341 & 0.343$\pm$0.221 \\
Testing & 0.474$\pm$0.272 & - & 0.474$\pm$0.272 & - & - & 0.608$\pm$0.108 & 0.407$\pm$0.302 \\
\hline
\end{tabular}%
\label{tab:categories_detach}%
\end{table}%
\end{document}
\begin{thebibliography}{10}
\urlstyle{rm}
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{\texttt{#1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
\expandafter\ifx\csname doiprefix\endcsname\relax\def\doiprefix{DOI: }\fi
\providecommand{\bibinfo}[2]{#2}
\providecommand{\eprint}[2][]{\url{#2}}
\bibitem{sankaridurg2021imi}
\bibinfo{author}{Sankaridurg, P.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Imi impact of myopia}}.
\newblock {\emph{\JournalTitle{Investigative Ophthalmology \& Visual Science}}}
\textbf{\bibinfo{volume}{62}}, \bibinfo{pages}{2--2} (\bibinfo{year}{2021}).
\bibitem{percival1987redefinition}
\bibinfo{author}{Percival, S.}
\newblock \bibinfo{journal}{\bibinfo{title}{Redefinition of high myopia: the
relationship of axial length measurement to myopic pathology and its
relevance to cataract surgery.}}
\newblock {\emph{\JournalTitle{Developments in ophthalmology}}}
\textbf{\bibinfo{volume}{14}}, \bibinfo{pages}{42--46}
(\bibinfo{year}{1987}).
\bibitem{ohno2021imi}
\bibinfo{author}{Ohno-Matsui, K.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Imi pathologic myopia}}.
\newblock {\emph{\JournalTitle{Investigative Ophthalmology \& Visual Science}}}
\textbf{\bibinfo{volume}{62}}, \bibinfo{pages}{5--5} (\bibinfo{year}{2021}).
\bibitem{vingolo2019pathologic}
\bibinfo{author}{Vingolo, E.~M.}, \bibinfo{author}{Napolitano, G.} \&
\bibinfo{author}{Casillo, L.}
\newblock \bibinfo{title}{Pathologic myopia: complications and visual
rehabilitation}.
\newblock In \emph{\bibinfo{booktitle}{Intraocular Lens}}, \bibinfo{pages}{67}
(\bibinfo{publisher}{IntechOpen}, \bibinfo{year}{2019}).
\bibitem{orlando2020refuge}
\bibinfo{author}{Orlando, J.~I.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Refuge challenge: A unified
framework for evaluating automated methods for glaucoma assessment from
fundus photographs}}.
\newblock {\emph{\JournalTitle{Medical image analysis}}}
\textbf{\bibinfo{volume}{59}}, \bibinfo{pages}{101570}
(\bibinfo{year}{2020}).
\bibitem{li2021applications}
\bibinfo{author}{Li, T.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Applications of deep learning in
fundus images: A review}}.
\newblock {\emph{\JournalTitle{Medical Image Analysis}}}
\textbf{\bibinfo{volume}{69}}, \bibinfo{pages}{101971}
(\bibinfo{year}{2021}).
\bibitem{hagiwara2018computer}
\bibinfo{author}{Hagiwara, Y.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Computer-aided diagnosis of
glaucoma using fundus images: A review}}.
\newblock {\emph{\JournalTitle{Computer methods and programs in biomedicine}}}
\textbf{\bibinfo{volume}{165}}, \bibinfo{pages}{1--12}
(\bibinfo{year}{2018}).
\bibitem{sengupta2020ophthalmic}
\bibinfo{author}{Sengupta, S.}, \bibinfo{author}{Singh, A.},
\bibinfo{author}{Leopold, H.~A.}, \bibinfo{author}{Gulati, T.} \&
\bibinfo{author}{Lakshminarayanan, V.}
\newblock \bibinfo{journal}{\bibinfo{title}{Ophthalmic diagnosis using deep
learning with fundus images--a critical review}}.
\newblock {\emph{\JournalTitle{Artificial Intelligence in Medicine}}}
\textbf{\bibinfo{volume}{102}}, \bibinfo{pages}{101758}
(\bibinfo{year}{2020}).
\bibitem{fu2019palm}
\bibinfo{author}{Fu, H.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Palm: Pathologic myopia
challenge}}.
\newblock {\emph{\JournalTitle{IEEE Dataport}}} (\bibinfo{year}{2019}).
\bibitem{baskaran2015prevalence}
\bibinfo{author}{Baskaran, M.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{The prevalence and types of
glaucoma in an urban chinese population: the singapore chinese eye study}}.
\newblock {\emph{\JournalTitle{JAMA ophthalmology}}}
\textbf{\bibinfo{volume}{133}}, \bibinfo{pages}{874--880}
(\bibinfo{year}{2015}).
\bibitem{ODIR-5K}
\bibinfo{title}{Ocular disease intelligent recognition odir-5k}.
\newblock \bibinfo{howpublished}{\url{https://odir2019.grand-challenge.org/}}.
\bibitem{de_vente_coen_2021_5745834}
\bibinfo{author}{de~Vente, C.} \emph{et~al.}
\newblock \bibinfo{title}{Rotterdam eyepacs airogs train set - part 2/2},
\url{10.5281/zenodo.5745834} (\bibinfo{year}{2021}).
\bibitem{biswas2022color}
\bibinfo{author}{Biswas, S.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Which color channel is better for
diagnosing retinal diseases automatically in color fundus photographs?}}
\newblock {\emph{\JournalTitle{Life}}} \textbf{\bibinfo{volume}{12}},
\bibinfo{pages}{973} (\bibinfo{year}{2022}).
\bibitem{app11020591}
\bibinfo{author}{Son, J.}, \bibinfo{author}{Kim, J.}, \bibinfo{author}{Kong,
S.~T.} \& \bibinfo{author}{Jung, K.-H.}
\newblock \bibinfo{journal}{\bibinfo{title}{Leveraging the generalization
ability of deep convolutional neural networks for improving classifiers for
color fundus photographs}}.
\newblock {\emph{\JournalTitle{Applied Sciences}}}
\textbf{\bibinfo{volume}{11}}, \url{10.3390/app11020591}
(\bibinfo{year}{2021}).
\bibitem{cui2021pathological}
\bibinfo{author}{Cui, J.}, \bibinfo{author}{Zhang, X.}, \bibinfo{author}{Xiong,
F.} \& \bibinfo{author}{Chen, C.-L.}
\newblock \bibinfo{journal}{\bibinfo{title}{Pathological myopia image
recognition strategy based on data augmentation and model fusion}}.
\newblock {\emph{\JournalTitle{Journal of Healthcare Engineering}}}
\textbf{\bibinfo{volume}{2021}} (\bibinfo{year}{2021}).
\bibitem{rauf2021automatic}
\bibinfo{author}{Rauf, N.}, \bibinfo{author}{Gilani, S.~O.} \&
\bibinfo{author}{Waris, A.}
\newblock \bibinfo{journal}{\bibinfo{title}{Automatic detection of pathological
myopia using machine learning}}.
\newblock {\emph{\JournalTitle{Scientific Reports}}}
\textbf{\bibinfo{volume}{11}}, \bibinfo{pages}{1--9} (\bibinfo{year}{2021}).
\bibitem{hemelings2021pathological}
\bibinfo{author}{Hemelings, R.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Pathological myopia classification
with simultaneous lesion segmentation using deep learning}}.
\newblock {\emph{\JournalTitle{Computer Methods and Programs in Biomedicine}}}
\textbf{\bibinfo{volume}{199}}, \bibinfo{pages}{105920}
(\bibinfo{year}{2021}).
\bibitem{guo2020lesion}
\bibinfo{author}{Guo, Y.} \emph{et~al.}
\newblock \bibinfo{title}{Lesion-aware segmentation network for atrophy and
detachment of pathological myopia on fundus images}.
\newblock In \emph{\bibinfo{booktitle}{2020 IEEE 17th international symposium
on biomedical imaging (ISBI)}}, \bibinfo{pages}{1242--1245}
(\bibinfo{organization}{IEEE}, \bibinfo{year}{2020}).
\bibitem{9193942}
\bibinfo{author}{Xie, R.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{End-to-end fovea localisation in
colour fundus images with a hierarchical deep regression network}}.
\newblock {\emph{\JournalTitle{IEEE Transactions on Medical Imaging}}}
\textbf{\bibinfo{volume}{40}}, \bibinfo{pages}{116--128},
\url{10.1109/TMI.2020.3023254} (\bibinfo{year}{2021}).
\bibitem{du2021deep}
\bibinfo{author}{Du, R.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Deep learning approach for
automated detection of myopic maculopathy and pathologic myopia in fundus
images}}.
\newblock {\emph{\JournalTitle{Ophthalmology Retina}}}
\textbf{\bibinfo{volume}{5}}, \bibinfo{pages}{1235--1244}
(\bibinfo{year}{2021}).
\bibitem{flitcroft2019imi}
\bibinfo{author}{Flitcroft, D.~I.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Imi--defining and classifying
myopia: a proposed set of standards for clinical and epidemiologic studies}}.
\newblock {\emph{\JournalTitle{Investigative ophthalmology \& visual science}}}
\textbf{\bibinfo{volume}{60}}, \bibinfo{pages}{M20--M30}
(\bibinfo{year}{2019}).
\bibitem{lin2020sustech}
\bibinfo{author}{Lin, L.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{The sustech-sysu dataset for
automated exudate detection and diabetic retinopathy grading}}.
\newblock {\emph{\JournalTitle{Scientific Data}}} \textbf{\bibinfo{volume}{7}},
\bibinfo{pages}{1--10} (\bibinfo{year}{2020}).
\bibitem{fu2019evaluation}
\bibinfo{author}{Fu, H.} \emph{et~al.}
\newblock \bibinfo{title}{Evaluation of retinal image quality assessment
networks in different color-spaces}.
\newblock In \emph{\bibinfo{booktitle}{International Conference on Medical
Image Computing and Computer-Assisted Intervention}}, \bibinfo{pages}{48--56}
(\bibinfo{organization}{Springer}, \bibinfo{year}{2019}).
\bibitem{raj2019fundus}
\bibinfo{author}{Raj, A.}, \bibinfo{author}{Tiwari, A.~K.} \&
\bibinfo{author}{Martini, M.~G.}
\newblock \bibinfo{journal}{\bibinfo{title}{Fundus image quality assessment:
survey, challenges, and future scope}}.
\newblock {\emph{\JournalTitle{IET Image Processing}}}
\textbf{\bibinfo{volume}{13}}, \bibinfo{pages}{1211--1224}
(\bibinfo{year}{2019}).
\bibitem{shao2017automated}
\bibinfo{author}{Shao, F.}, \bibinfo{author}{Yang, Y.}, \bibinfo{author}{Jiang,
Q.}, \bibinfo{author}{Jiang, G.} \& \bibinfo{author}{Ho, Y.-S.}
\newblock \bibinfo{journal}{\bibinfo{title}{Automated quality assessment of
fundus images via analysis of illumination, naturalness and structure}}.
\newblock {\emph{\JournalTitle{IEEE Access}}} \textbf{\bibinfo{volume}{6}},
\bibinfo{pages}{806--817} (\bibinfo{year}{2017}).
\bibitem{fleming2006automated}
\bibinfo{author}{Fleming, A.~D.}, \bibinfo{author}{Philip, S.},
\bibinfo{author}{Goatman, K.~A.}, \bibinfo{author}{Olson, J.~A.} \&
\bibinfo{author}{Sharp, P.~F.}
\newblock \bibinfo{journal}{\bibinfo{title}{Automated assessment of diabetic
retinal image quality based on clarity and field definition}}.
\newblock {\emph{\JournalTitle{Investigative ophthalmology \& visual science}}}
\textbf{\bibinfo{volume}{47}}, \bibinfo{pages}{1120--1125}
(\bibinfo{year}{2006}).
\bibitem{niemeijer2006image}
\bibinfo{author}{Niemeijer, M.}, \bibinfo{author}{Abramoff, M.~D.} \&
\bibinfo{author}{van Ginneken, B.}
\newblock \bibinfo{journal}{\bibinfo{title}{Image structure clustering for
image quality verification of color retina images in diabetic retinopathy
screening}}.
\newblock {\emph{\JournalTitle{Medical image analysis}}}
\textbf{\bibinfo{volume}{10}}, \bibinfo{pages}{888--898}
(\bibinfo{year}{2006}).
\bibitem{fang2022dataset}
\bibinfo{author}{Fang, H.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Dataset and evaluation algorithm
design for goals challenge}}.
\newblock {\emph{\JournalTitle{arXiv preprint arXiv:2207.14447}}}
(\bibinfo{year}{2022}).
\bibitem{ronneberger2015u}
\bibinfo{author}{Ronneberger, O.}, \bibinfo{author}{Fischer, P.} \&
\bibinfo{author}{Brox, T.}
\newblock \bibinfo{title}{U-net: Convolutional networks for biomedical image
segmentation}.
\newblock In \emph{\bibinfo{booktitle}{International Conference on Medical
image computing and computer-assisted intervention}},
\bibinfo{pages}{234--241} (\bibinfo{organization}{Springer},
\bibinfo{year}{2015}).
\bibitem{he2016deep}
\bibinfo{author}{He, K.}, \bibinfo{author}{Zhang, X.}, \bibinfo{author}{Ren,
S.} \& \bibinfo{author}{Sun, J.}
\newblock \bibinfo{title}{Deep residual learning for image recognition}.
\newblock In \emph{\bibinfo{booktitle}{Proceedings of the IEEE conference on
computer vision and pattern recognition}}, \bibinfo{pages}{770--778}
(\bibinfo{year}{2016}).
\bibitem{fang2022adam}
\bibinfo{author}{Fang, H.} \emph{et~al.}
\newblock \bibinfo{journal}{\bibinfo{title}{Adam challenge: Detecting
age-related macular degeneration from fundus images}}.
\newblock {\emph{\JournalTitle{IEEE Transactions on Medical Imaging}}}
(\bibinfo{year}{2022}).
\end{thebibliography}