Research Progress in Artificial Intelligence for Central Serous Chorioretinopathy: A Systematic Review

Abstract

Introduction

This review synthesizes advancements in artificial intelligence (AI) applications for central serous chorioretinopathy (CSCR), analyzing challenges and outlining future research directions to guide personalized diagnostic and therapeutic strategies.

Methods

A systematic literature search was conducted in the Web of Science database using a comprehensive AI-related keyword set (e.g., “deep learning,” “neural networks,” “computer vision”) combined with “central serous chorioretinopathy.” The search yielded 698 records, with 73 original research studies selected after excluding reviews and non-empirical work based on predefined criteria.

Results

The application of AI in CSCR has progressed from disease classification to dynamic prognostic prediction, leveraging multimodal data fusion (e.g., optical coherence tomography [OCT] and OCT angiography [OCTA] and fundus fluorescein angiography [FFA]) to enhance both qualitative and quantitative diagnostic accuracy. AI models outperform clinical experts in classifying retinal disease subtypes and segmenting lesions. However, clinical translation faces infrastructural barriers (e.g., incompatible PACS systems) and limited physician trust due to “black box” decision-making. New approaches, such as explainable AI (XAI), are being integrated to enhance the transparency and clinical applicability of AI models. Key limitations involve single-center data dependency, interobserver annotation variability, and the inability of static frameworks to capture dynamic lesion progression.

Conclusion

AI enhances CSCR diagnosis and subtyping efficiency. To optimize clinical translation, future research should focus on multicenter data integration, dynamic visualization frameworks, and standardized guidelines, promoting interdisciplinary collaboration and prospective trials for personalized treatment strategies. Incorporating federated learning for privacy-preserving data sharing and prioritizing explainability will be essential to overcoming barriers to physician adoption and improving trust in AI-driven clinical decision-making. Future efforts should focus on creating dynamic systems that provide real-time insights into lesion progression and integrating them with standardized protocols for wider clinical adoption.

Key Summary Points

AI significantly enhances diagnostic accuracy and efficiency in CSCR, outperforming clinicians in classifying disease subtypes and segmenting key lesions like subretinal fluid using OCT and fundus imaging.

Multimodal data fusion (e.g., OCT, OCTA, FFA, clinical parameters) is crucial for improving AI performance in diagnosis, lesion quantification, and predicting treatment response and recurrence.

Current AI models face limitations in generalizability and clinical translation due to reliance on single-center data, annotation variability, static frameworks lacking dynamic monitoring, and the absence of standardized CSCR classification and treatment protocols.

Future research must prioritize multicenter data integration, dynamic visualization frameworks, standardized guidelines, and prospective clinical trials to optimize AI's clinical utility and enable personalized CSCR management.

Introduction

Central serous chorioretinopathy (CSCR) is an idiopathic macular disorder characterized by retinal pigment epithelium (RPE) dysfunction and choroidal hyperpermeability, leading to neurosensory retinal detachment (NRD) and/or pigment epithelium detachment (PED) (Fig. 1). NRD manifests as subretinal fluid accumulation in the macular region, while PED reflects structural separation of the RPE–Bruch membrane complex. These pathoanatomical features correlate with visual disturbances and serve as critical biomarkers for diagnosis and staging [1, 2]. Chronic CSCR (cCSCR), characterized by persistent choroidal abnormalities such as vascular dilation and leakage, can lead to irreversible photoreceptor damage. Recent research further highlights the pivotal role of choroidal vascular dynamics in CSCR progression [3, 4]. Artificial Intelligence (AI), especially machine learning (ML), has revolutionized CSCR research by enhancing diagnostic efficiency and accuracy, enabling precise analysis of complex medical images [5–8], and driving the standardization of clinical research evaluation frameworks through methodological innovations [9]. AI models, trained on multimodal data from optical coherence tomography (OCT) and color fundus photography (CFP), have shown superior performance in lesion detection and quantification [10, 11]. This review discusses AI applications in CSCR diagnosis, segmentation, and prognosis prediction, examining current challenges and future directions to support clinical practice.

Fig. 1.

a Horizontal structural optical coherence tomography (OCT) scan passing through the neurosensory retinal detachment (NRD) and pigment epithelium detachment (PED); NRD manifests as subretinal fluid accumulation in the macular region, while PED reflects structural separation of the RPE–Bruch membrane complex. b CSCR is defined by fluorescein fundus angiography (FFA) demonstrating leakage. c Scanning laser ophthalmoscope (SLO) image of CSCR-affected right eye. d Color fundus photography (CFP) of CSCR-affected right eye

Methods

In this review, we performed a comprehensive search within the Web of Science using a combination of keywords including “artificial intelligence,” “machine learning,” “neural networks,” “deep learning,” “transfer learning,” “supervised learning,” “unsupervised learning,” “computer vision systems,” “ensemble learning,” “reinforcement learning,” “evolutionary computation,” “large language models,” and “central serous chorioretinopathy.” The objective was to investigate the extensive applications of AI in the context of CSCR. To ensure the relevance and quality of the findings, we excluded review articles, commentaries, conference abstracts, and unpublished research from the analysis. A detailed examination of the 73 included articles was conducted. Figure 2 illustrates the workflow of literature search and screening. This review systematically examines four pivotal applications of AI in CSCR research: diagnostic classification, lesion segmentation, quantitative lesion measurement, and prognosis/recurrence prediction, while integrating real-world intelligent platform research to establish a holistic overview of the field.

Fig. 2.

Literature search flowchart

This article is based on previously conducted research and does not contain any new research with human participants or animals performed by any of the authors.

Analysis of AI Applications in the Diagnosis of CSCR

This section summarizes research on AI models for the automatic detection of CSCR and differentiation between acute and chronic cases using CFP and OCT images. For diagnosis-specific details of each article, refer to Table 1.

Table 1.

AI in CSCR diagnosis and segmentation

Category	References	Datasets		AI-related methods		Performance	Clinical context	Source
Diagnosis	Zhen et al. [12]	2504 CFP		DL architecture called Inception-V3		AUC-ROC 0.934; kappa coefficients 0.59 (vs. rater 1), 0.33 (vs. rater 2)	Presence or absence of CSCR	Retina–The Journal of Retinal and Vitreous Diseases
Diagnosis	Anitha et al. [13]	420 retinal images, including CSCR		ANN-based system for pathology detection		Accuracy 97.7%; specificity 98%; sensitivity 96%	Classifying retinal image abnormalities with CSCR included	British Journal of Ophthalmology
Diagnosis	Wang et al. [14]	16,000 retinal images (2k CSCR)		Light-weight auxiliary models by transfer learning		MobileNetV3 model to diagnose CSCR AUC was 0.803, ShufflenetV2 model was 0.844	Common retinal disease classification	Graefes Archive for Clinical and Experimental Ophthalmology
Diagnosis	Komuku et al. [18]	400 CFPs		VGG-16 deep convolutional neural network		Correlation of choroidal thickness: 0.68 (normal) vs. 0.48 (CSCR)	CSCR vs. healthy controls	Scientific Reports
Diagnosis	Shoba et al. [19]	150 OCT retinal images		Shallow neural network with Powell-Beale		Accuracy 93.3%; precision 93.3%; specificity 93.5%; sensitivity 94.1%	CSCR vs. healthy controls	Computer Systems Science and Engineering
Diagnosis	Xu et al. [20]	358 OCT B-scan images		DWT, LBP, and multi-instance learning (MIL), with MIL as classifier		K = 35, achieving 99.58% accuracy; K = 40, accuracy was 100%	CSCR vs. healthy controls	Computers in Biology and Medicine
Diagnosis	Yoon et al. [21]	2360 SD-OCT images from 220 participants		Deep learning systems (DLS) model		CSCR diagnosis: accuracy 93.8%; sensitivity 90.0%; specificity 99.1%; AUROC 98.9%; chronic vs. acute distinction: accuracy 97.6%; sensitivity 100%, specificity 92.6%; AUROC 99.4%	Diagnosing CSCR and distinguishing between chronic CSCR and acute CSCR	Scientific Reports
Diagnosis	Yoon et al. [22]	1693 SD-OCT images		Based on CNN architecture VGG-16		AUROC 99.5%; accuracy 96.3%; sensitivity 97.1%; specificity 95.7%	Chronic vs. acute CSCR distinction	Journal of Medical Internet Research
Diagnosis	Yoon et al. [23]	3209 SD-OCT images from 435 patients with CSCR		Based on the CNN architecture, VGG-16, built a DL model		Average cross-validation accuracy 70.0%; accuracy in independent dataset 76.8%	Classify CSCR stages: acute, non-resolving, chronic, inactive	Scientific Reports
Diagnosis	Ko et al. [24]	7425 SD-OCT images from 297 participants		DL model based on SIP and FD classifiers		Sensitivity 94.9%; accuracy 94.2%; specificity 99.1%	Identify normal, acute, and chronic CSCR	Scientific Reports
Diagnosis	Üzen et al. [25]	2464 OCT images (1201 acute CSCR, 343 chronic CSCR, 920 healthy)		Custom multilevel CNN with groupwise and pointwise convolutions		Accuracy 98.42%; precision 97.18%; recall 97.93%; F1 score 97.52	Automated classification of acute vs. chronic CSCR	Optics and Laser Technology
Diagnosis	Hassan et al. [11]	Enhanced dataset: 580 OCT and 1220 retinal images		DL classifiers: DenseNet and DarkNet		Accuracy 99.78%; sensitivity 99.6%; specificity 100%; F1 score 99.52%	CSCR vs. healthy controls	Multimedia Tools and Applications
Diagnosis	Aoyama et al. [31]	100 en face OCT images from 100 participants		Two CNNs: Neural Network Console (NNC; SONY) and Keras-TensorFlow (Google)		Mean accuracy rate was 88% for Keras and 95% for NNC	CSCR vs. healthy controls	PLoS One
Diagnosis	Gan et al. [26]	1213 ME SD-OCT images (306 CSCR)		AI classifiers: support vector machine (SVM)		Accuracy 93.8%; CSC groups AUC 100%	Differentiating DME/AME/RVO/CSCR via SD-OCT	Frontiers in Neuroscience
Diagnosis	Han et al. [27]	6063 SD-OCT images (521 patients; 191 CSCR, 47 healthy)		DL model based on CNN		Accuracy 99.7%; chronic CSCR (F1 0.97); acute (0.94)	nAMD vs. CSCR subtypes via SD-OCT	Journal of Clinical Medicine
Diagnosis	Khan et al. [28]	2998 OCT images, including CSCR cases		Combine transfer learning and ant colony optimization		Deep features and hybrid selection: 99.1% accuracy (97.4% without ACO)	Classify retinal abnormalities (including CSCR)	Sensors
Diagnosis	Kang et al. [29]	1898 OCT B-scans, including CSCR cases		Pre-trained CNN-based DL models		Xception model: accuracy/F1 score 92%	Classifying thick vs. non-thick choroid diseases	Graefes Archive for Clinical and Experimental Ophthalmology
Diagnosis	Mirshahi et al. [17]	82 patients with CSCR: EDI-OCT imaging		An unsupervised learning technique: cluster analysis		Sensitivity 1.000; specificity 0.986	Thickened choroidal diseases (incl. CSCR) vs. healthy	Scientific Reports
Diagnosis	Lee et al. [30]	1267 non-thick and 977 thick choroid images		Deep convolutional neural networks (DCNN)		Accuracy 94.7% (validation set)/90.1% (test set); F1 score 0.904	Classifying thick vs. non-thick choroid diseases	Translational Vision Science & Technology
Segmentation	Wu et al. [36]	31 retinal SD-OCT volumes		Thickness map prior, spatial smoothness		TPVF 94.3%, FPVF 0.97%, PPV 93.6%	Segmentation of SRF associated with CSCR	IEEE Transactions on Biomedical Engineering
Segmentation	Bekalo et al. [37]	20 SD-OCT cubes from patients with NRD		k-means algorithm, graph search and graph cut techniques		Compared with the two experts, average DSC were 90.78% and 92.04%	Demarcation of SRF associated with NRD	IEEE Access
Segmentation	Xu et al. [35]	12,532 retinal photographs		Cascaded DL system		AUC 0.910; sensitivity 92.3%; specificity 82.0%	SRF detection and referral recommendations	Frontiers in Bioengineering and Biotechnology
Segmentation	Pawan et al. [39]	25 retinal OCT volumes		DRIP-Caps: modified SegCaps (Capsule Network-based)		Precision 94.17; recall 94.06; DSC 94.04	Segmentation of SRF associated with CSCR	Medical & Biological Engineering & Computing
Segmentation	Yoo et al. [34]	194 CSCR retinal images (SRF lesions)		U-Net and cGAN (pix2pix)		Jaccard index 0.619, DSC 0.763	Segmentation of SRF associated with CSCR	Translational Vision Science & Technology
Segmentation	Goyanes et al. [40]	124 3D retinal OCT volumes		An advanced DL architecture: nnU-Net		Precision 0.827 ± 0.256; recall 0.734 ± 0.228; accuracy 0.998 ± 0.003; Jaccard 0.644 ± 0.253; DSC 0.745 ± 0.249	Segmentation of liquid region associated with CSCR	Journal of Imaging Informatics in Medicine
Segmentation	Gende et al. [38]	557 OCT images from participants		Two backbone segmentation architectures: FPN and U-Net		Dice 0.868 ± 0.056; kappa 0.960 (experts 0.951)	Segmenting fluid regions (CSCR-associated)	Journal of Imaging Informatics in Medicine
Segmentation	Xing et al. [48]	23 SD-OCT volumes from 12 patients		Located-CNN model based on weakly supervised learning		Classification accuracy 97.25%; localization precision 90.25%	Identify the location of SRD in the CSCR	Biomedical Optics Express
Segmentation	Chen et al. [42]	2104 FFA images from 262 patients		DL algorithm (DLA)		Accuracy 93.4%; Dice score 0.949	Segmentation of CSCR leak points in FFA	Graefes Archive for Clinical and Experimental Ophthalmology
Segmentation	Maruyama et al. [43]	34 patients with CSCR		DL analysis		Choroidal volume: AUC 0.84; sensitivity 69.57%; specificity 85.71%	3D segmentation of choroidal blood vessels	Translational Vision Science & Technology
Segmentation	Zhu et al. [44]	Enrolled 17 patients with CSCR		Segmentation model based on ML		Dice score 77.28	Segmentation of choroidal blood vessels	Health Data Science
Segmentation	Liu et al. [45]	138,000 OCT images (multidevice)		Multitask convolutional neural network model		DSC 0.956	3D segmentation of retinal blood vessels	Biomedical Optics Express
Segmentation	Xu et al. [46]	20 CSCR retinal images (test set)		A few-shot learning-based method		Best accuracy 96.96%; sensitivity 83.44%; specificity 98.05%; AUC 98.03%	Segmentation of retinal blood vessels	Frontiers in Medicine
Segmentation	Wu et al. [47]	37 SD-OCT volumes from patients with CSCR		Construct a probability graph using random forest classification		For NRD segmentation: TPVF 92.1%; FPVF 0.53%; PPV 94.7%; DSC 93.3%; for PED segmentation: TPVF 92.5%; FPVF 0.14%; PPV 80.9%; DSC 84.6%	Segmentation of NRD and PED associated with CSCR	Biomedical Optics Express
Segmentation	Gao et al. [41]	52 SD-OCT volumes from 35 patients		DA-FCN: double-branched area-constrained fully convolutional network		TPVF 94.3; Dice 95.3; PPV 96.4	Segmentation of NRD and PED associated with CSCR	Computer Methods and Programs in Biomedicine
Segmentation	Gao et al. [49]	10 SD-OCT cubes, a total of 200 B-scan images		Graph search algorithm		Boundary positioning difference 3.68 ± 2.96 μm; thickness difference 5.84 ± 4.78 μm	Retinal boundary segmentation refines CSCR lesion extent	Multimedia Tools and Applications
Segmentation	Xue et al. [50]	1280 OCT images		SNP systems and CNNs		Processing time 0.04 s; Dice ratios: 0.93 ± 0.04 and 0.95 ± 0.02	Segmenting CSCR lesions and EZ-BM	Information Fusion
Segmentation	Wei et al. [51]	1000 OCT images from 100 patients with CSCR		Lightweight DBR neural network		Accuracy 75.38%; DSC 82.20%; recall rate 74.37%	Segmentation of HF associated with CSCR	Journal of Digital Imaging
Segmentation	Santarossa et al. [52]	326 patients with CSCR		Ensemble model: multiple U-Net (identical architectures)		Dice scores were 64–81% for HF and 21–41% for RA	HF and RA segmentation in CSCR-related FAF	Journal of Imaging

AUC-ROC area under the receiver operating characteristic curve, AUC area under the curve, DSC Dice similarity coefficient, DWT discrete wavelet transform, LBP local binary patterns, SIP single-image prediction, FD final decision, FPVF false positive volume fraction, TPVF true positive volume fraction, PPV positive predictive value, cGAN conditional generative adversarial network, FPN feature pyramid network, SNP systems spiking neural P systems

Automated Diagnosis of CSCR Based on CFP

Zhen et al. [12] employed an Inception-V3 DL architecture for binary classification of CSCR, achieving an AUROC of 0.934. However, the kappa coefficients between this model and the two graders were unsatisfactory, at only 0.59 and 0.33, respectively. In classifying common fundus diseases, Anitha et al. [13] utilized artificial neural network (ANN) technology, achieving an average accuracy of 97.7% in identifying CSCR cases within a dataset of 420 retinal images. To reduce the time and scale burdens of classical classification models, Wang et al. [14] applied transfer learning techniques, resulting in a lightweight auxiliary model with classification accuracy exceeding 80%.

CSCR is often associated with abnormal choroidal thickness [15, 16]. Mirshahi et al. [17] employed an unsupervised clustering analysis method to identify key parameters from EDI-OCT images. They found that a choroidal thickness (CT) of 389 µm and a Haller ratio (HR) of 0.79 were able to distinguish thickened from normal choroid with nearly 100% sensitivity and specificity. Komuku et al. [18] achieved automatic classification and estimation of thickened choroidal diseases using CFP based on DL techniques; however, this article is limited by a small dataset and strong dependency on image quality.

Automated Diagnosis Based on OCT

Automated diagnosis using OCT leverages its high-resolution, 3D imaging capabilities to visualize retinal and choroidal structures, facilitating accurate CSCR detection.

Shoba et al. [19] introduced a shallow neural network with the Powell-Beale (SNNWPB) that classifies OCT images as normal or CSCR with 93.3% accuracy after preprocessing and feature extraction. Xu et al. [20] developed a method that extracts texture features from OCT B-scans without segmentation, employing a multifusion threshold decision mechanism for efficient classification.

Classifying CSCR into acute, chronic, non-resolving and inactive is vital for timely intervention. Yoon et al. [21] evaluated a deep learning system (DLS) that diagnoses spectral domain optical coherence tomography (SD-OCT) images, achieving 93.8% accuracy without segmentation. A convolutional neural networks (CNN)-based VGG-16 architecture later improved this to 96.3% [22]. This system identifies CSCR subtypes, although averaging 70% cross-validation accuracy, it still outperformed expert clinicians [23].

In complex clinical settings, diagnosis often involves multiple OCT images. Ko et al. [24] synthesized several images for cCSCR diagnosis, achieving 94.2% accuracy. The multilayer feature fusion CNN for OCT image analysis, recently proposed by Üzen et al. [25], generates more representative feature maps by extracting and fusing features from different hierarchical levels. This model demonstrated superior performance in differentiating between acute and chronic CSCR, achieving 97% accuracy, recall, and F1 score. Hassan et al. [11] utilized a public database to enhance a DenseNet classifier, reaching over 98% accuracy. Gan et al. [26] achieved 100% accuracy by integrating omics data with SD-OCT features, showcasing the potential of multimodal data fusion. Han et al. [27] attained 91.1% accuracy in subtype classification using transfer learning and hybrid data augmentation, while Khan et al. [28] combined transfer learning with an ant colony optimization algorithm for 99.1% accuracy in detecting retinal diseases, including CSCR.

Kang et al. [29] differentiated between thick and non-thick choroidal diseases in OCT B-scans, and Lee et al. [30] employed deep convolutional neural networks (DCNNs) to analyze 3D data from en face OCT images, achieving 94.7% accuracy which furthered the progress of automated choroid analysis. Investigating the choroidal vasculature could provide perspectives for CSCR diagnosis. Aoyama et al. [31] analyzed choroidal vessels in en face OCT images for differentiating CSCR-affected eyes from normal ones, although this was a retrospective study with limited cases. This evolution of techniques reflects a trend towards more sophisticated models and multimodal approaches, enhancing the accuracy of CSCR diagnosis in clinical practice.

The integration of OCT’s superior structural resolution with CFP’s wide-field visualization enables reliable diagnosis of CSCR and effectively differentiates confusable vascular pathologies (e.g., diabetic retinopathy, wet age-related macular degeneration) [32]. Viggiano et al. [33] demonstrated that AI models significantly enhance OCT scan interpretation by enabling precise identification of CSCR-critical features, such as subretinal fluid (SRF) and choroidal thickness, ultimately underscoring AI’s pivotal role in detecting choroidal abnormalities to improve diagnostic accuracy.

AI has significantly improved the diagnostic accuracy of CSCR. The next critical steps involve precise lesion localization and quantitative assessment.

Research Analysis on the Application of AI in CSCR Segmentation

By employing automated segmentation of multimodal imaging data, researchers accurately localized and analyzed key CSCR features, such as SRF, leakage points (LP), and vascular abnormalities, thus advancing quantitative CSCR diagnosis. For segmentation-specific details of each article, refer to Table 1.

AI for SRF Segmentation

To segment SRF regions in CFP for CSCR detection, various DL approaches have been developed. Yoo et al. [34] proposed a U-Net model with the Pix2Pix algorithm, achieving Jaccard index and Dice similarity coefficient (DSC) of 0.619 and 0.763 in a 194-image dataset, though accuracy for atypical SRF cases needs improvement. Xu et al. [35] created a cascaded DL system for SRF detection and macular status identification, but true fluid segmentation remains unachieved. Wu et al. [36] propose an OCT en face fundus-driven SRF segmentation approach for CSCR, effectively addressing the segmentation of retinal fluid within the boundaries of relevant retinal tissues. Bekalo et al. [37] advanced a 3D segmentation technique using graph search and cut, achieving over 90% DSC for NRD-related SRF, though NRD fluid presence may hinder layer segmentation accuracy. Gende et al. [38] demonstrated that their DL models provided more consistent SRF segmentation than expert agreement, but relied on a large dataset and numerous parameters. SegCaps and DRIP-Caps, introduced by Pawan et al. [39], based on the capsule network concept, were designed to simplify network complexity while maintaining performance. However, the evaluations relied on single-expert annotations, which risks introducing bias. Goyanes et al. [40] employed a 3D CNN for CSCR fluid segmentation, achieving 93% accuracy with 2769 OCT images, while Gao et al. [41] developed a dual-branch, region-constrained fully convolutional network (DA-FCN), achieving a DSC of 95.30% on 23 OCT volumes without prior retinal layer segmentation. Collectively, these methods indicate that enhancing segmentation robustness through diverse technical approaches—ranging from complex architectures to lightweight designs—represents a key objective in addressing clinical needs.

AI for LP Segmentation

Chen et al. [42] developed an attention-gated network (AGN) for automated LP detection in FFA images, achieving 93.4% accuracy and a DSC of 0.949, offering potential guidance for laser therapy in CSCR.

AI for Choroidal and Retinal Vessel Segmentation in CSCR

DL-based segmentation methods offer crucial support for quantitative vascular assessment in the research of blood vessels. Maruyama et al. [43] employed DL for 3D choroidal visualization and quantitative analysis of vascular parameters. Furthermore, Zhu et al. [44] developed the first non-invasive, OCT angiography (OCTA)-like clinical tool for the detailed 3D reconstruction of OCT B-scans. Liu et al. [45] introduced a multitask convolutional neural network model, termed 3D OCTA-SCNet, for multimodal data segmentation, integrating OCT and OCTA. To alleviate the burden of data annotation, Xu et al. [46] employed an ensemble learning framework trained on public datasets to reduce the need for extensive data labeling and explored few-shot learning for retinal vessel segmentation. Further clinical validation of its robustness is necessary.

AI for Segmentation of NRD and PED in CSCR

Wu et al. [47] automated NRD/PED differentiation in SD-OCT images using a 3D random forest-based approach, achieving DSCs of 93.3% (NRD) and 84.6% (PED). Xing et al. [48] proposed a weakly supervised located-CNN model for NRD localization, attaining 97.25% classification accuracy with minimal annotation.

AI-Based Segmentation of Boundaries and Hyperreflective Foci (HF) in CSCR

Gao et al. [49] optimized the outer nuclear layer (ONL)-inner segment (IS) boundary segmentation using a graph search algorithm, achieving results comparable to manual methods. Concurrently, Xue et al. [50] integrated spiking neural P systems (SNP) with CNN to segment CSCR-related ellipsoid zone-Bruch’s membrane (EZ-BM) regions, attaining an average DSC exceed 90%.

Wei et al. [51] developed a lightweight DBR network for HF segmentation in OCT images, achieving accuracy, DSC, and recall of 75.38%, 82.20%, and 74.37%, respectively. However, limitations included dataset homogeneity. Santarossa et al. [52] applied ensemble learning to delineate HF and reduced autofluorescence (RA) in fundus autofluorescence (FAF) images, achieving DSCs of 64–81% (HF) and 21–41% (RA).

Research Analysis on the Application of AI in CSCR Measurement

In CSCR research, precise measurement of lesion diameter and SRF 3D morphology is crucial for assessing severity, optimizing treatment, and monitoring outcomes. Advances in image fusion, DL algorithms, and localization technologies have improved lesion quantification, aiding clinical decisions. The measurement section of Table 2 contains further details.

Table 2.

AI measurement, recurrence and prognosis prediction, and AI platform applications in CSCR

Category	References	Dataset	Core content of the article	Limitation analysis	Source
Measurement	Xu et al. [53]	Validation set: 396 angiography–CFP pairs	Improved DL method based on NSCT integrated multimodal retinal images for CSCR lesion localization	Fusion image quality needs improvement	Mathematical Biosciences and Engineering
Measurement	Xu et al. [54]	912 test images included patients with CSCR	Multitask ILM enables CSCR lesion boundary identification and area diameter calculation	PED reduces diameter measurement accuracy	Computers in Biology and Medicine
Measurement	Xu et al. [56]	Testing phase uses data from 25 patients, including 912 OCT B-scan images	Combined ILM with SARLM enhanced lesion localization for CSCR diameter measurement	Limited sample size; joint framework efficacy needs more validation and improvement	Computational Intelligence and Neuroscience
Measurement	Xu et al. [55]	912 OCT B-scan and SLO images used for model testing	DL method locates lesion boundaries in multimodal retinal images, enables automatic cross-modal operations via LBLM	Boundary localization depends on image quality	Computers in Biology and Medicine
Measurement	Xu et al. [57]	Conduct experiments on 1152 samples	AGT method optimizes SRF height measurement	In complex lesions and noise, these methods need optimization	Computers in Biology and Medicine
Measurement	Hara et al. [58]	58 patients with CSCR	3D choroidal vasculature visualization and quantitative analysis of vessels and stroma post PDT and PC in patients with CSCR	Limited sample size; a brief post-treatment follow-up period	Translational Vision Science & Technology
Recurrence prediction	Xu et al. [59]	461 patients with CSCR from two hospitals	ML model using clinical/imaging data to predict CSCR recurrence	Limited patient diversity and short follow-up	Frontiers in Physiology
Recurrence prediction	Seiler et al. [62]	Total of 255 patients with CSCR were enrolled	Multivariate logistic regression identified image biomarkers linked to CSCR recurrence	Complexity in SRF classification; reliance on expert intervention	Scientific Reports
Prognosis Prediction	Xu et al. [60]	416 patients with CSCR from two hospitals	ML predicts subretinal fluid absorption (SFA) post laser treatment	Relies on singular validation data	Annals of Translational Medicine
Prognosis prediction	Xu et al. [61]	416 patients with CSCR from two hospitals	Researchers utilized a GAN to achieve first prediction of post-treatment visual acuity and generation of OCT images	Short observation period; limited data volume	Frontiers in Bioengineering and Biotechnology
Prognosis prediction	Jee et al. [64]	Multiple OCT images of 832 patients with CSCR	Using multiple OCT image sets to evaluate performance of a DL model for CSCR prognosis	Model dependent on single OCT system; no standardized disease classification	Scientific Reports
Prognosis prediction	Pfau et al. [65]	SD-OCT images of 57 patients with cCSCR	DL infers post-treatment retinal sensitivity	Model limited to patients with cCSCR meeting trial criteria	Scientific Reports
Prognosis prediction	Arrigo et al. [68]	96 patients with CSCR	AI evaluates treatment response and MNV development	Small cohort; requires validation via controlled trials	Ophthalmology and Therapy
Prognosis prediction	Ra et al. [69]	440 patients with CSCR	Trained UNIT, CycleGAN and RegGAN to generate 3-month predictive OCT B-scans from baseline images	Single-center retrospective design; moderate sample size; requirement for multimodal data and longer follow-up periods	Graefes Archive for Clinical and Experimental Ophthalmology
Prognosis prediction	Fernandez-Vigo et al. [66]	175 patients with CSCR on half-fluence PDT	DL predicts PDT response in cCSCR	Single data source; multidimensional analysis needed	Photodiagnosis and Photodynamic Therapy
Prognosis prediction	Yoo et al. [67]	166 patients with cCSCR and 745 healthy control eyes	Two-stage DL predicts 1-year PDT outcome	Model–clinician consistency unverified; interpretability barriers	Scientific Reports
Prognosis prediction	López-Varela et al. [63]	52 patients (821 CC OCTA images)	Unsupervised CNN assesses choroidal flow changes	No comparative standards; limited validation data	Computerized Medical Imaging and Graphics
AI platform	Kim et al. [70]	284 UWF ICGA images (77 patients with CSCR)	Auto ML platform classifies thick choroid diseases via UWF ICGA	Improved performance with larger datasets; subjective classification; high-quality imaging essential	British Journal of Ophthalmology
AI platform	Desideri et al. [71]	OCT images of 144 patients	AI platform assesses OCT features to predict CSCR progression	Limited sample size	Retina–The Journal of Retinal and Vitreous Diseases
AI platform	Desideri et al. [72]	160 eyes from 144 patients with CSCR	AI platform identifies OCT biomarkers for CSCR diagnosis	Longitudinal validation required; critical biomarker thresholds undefined	International Journal of Retina and Vitreous
AI platform	Desideri et al. [73]	115 eyes from 101 patients with CSCR	AI platform predicts CSCR progression using baseline PED features	Requires further validation	Ophthalmologica

NSCT non-subsampled contourlet transform, ILM initial location module, SARLM single agent reinforcement learning module, LBLM lesion boundary location module, SLO scanning laser ophthalmoscope, AGT adaptive gradient threshold, GAN generative adversarial network, MNV macular neovascularization, CC OCTA choriocapillaris optical coherence tomography angiography

Multimodal Image Fusion and Intelligent Localization Techniques

Xu et al. [53] addressed limitations in traditional laser surgery by introducing a multimodal image fusion method, providing clear visualization of lesion structures for clinical use. Building on this, Xu et al. [54–56] advanced CSCR lesion analysis. They developed an iterative learning model (ILM-GRESS) [54] for rapid lesion boundary localization (0.1754 s) and an ILM-SARLM framework [56] achieving an 83.00% correction rate for accurate lesion diameter measurement. Additionally, a lesion boundary localization module (LBLM) [55] integrating CFP and OCT data improved the correction rate to 92.11%.

Implementation of Multidimensional Measurements of SRF in CSCR

To overcome the limitation of traditional SRF measurements being restricted to horizontal diameters, Xu et al. [57] proposed an integrated learning method that enhances key point location accuracy and optimizes SRF height measurement. Validation on 1152 samples showed effective performance in providing clear visualization of lesion structures, though further improvements are necessary for robustness against complex lesions and noise interference.

Three-Dimensional Volumetric Quantitative Analysis of Choroidal Vasculature and Stroma

Hara et al. [58] found that both photodynamic therapy (PDT) and laser photocoagulation (PC) significantly reduced choroidal vessel volume, with notable reductions in vascular and stromal volume, with notable reductions in vascular and stromal volume rates in the central 1.5-mm circle (PDT: P = 0.00029; PC: P = 0.0014) and with similar trends in the 1.5–6-mm torus area. The article highlights the ability to detect subtle structural changes and provides clearer visualization of lesions compared to conventional 2D methods.

Research Analysis on AI in CSCR Recurrence and Prognosis Prediction

Recurrence Prediction: From Feature Integration to Dynamic Analysis

Xu et al. [59] developed an ML model using 25 clinical parameters and 145 imaging features to predict CSCR recurrence. They found that retinal-choroidal features influenced short-term (3-month) recurrence, while lifestyle factors impacted long-term (6-month) outcomes. Later, they developed a predictive system [60] for SFA dynamics in 465 patients, identifying factors at 1-, 3-, and 6-month stages, and indicated that ML using multimodal imaging-derived data can reliably anticipate subretinal fluid resolution, reflecting the resolution of CSCR. Additionally, they utilized generative adversarial networks (GANs) [61] to predict visual acuity and generate synthetic OCT images. Seiler et al. [62] identified nine recurrence-related biomarkers, including choroidal vascularity index (CVI), and emphasized the role of their temporal evolution in prediction accuracy. López-Varela et al. [63] analyzed 821 OCTA images and proposed flow signal void changes as biomarkers for PDT response. They also developed a flow void segmentation method to quantify blood flow abnormalities.

Prognostic Assessment: From Structural Damage to Functional Prediction

Xu et al. [59] highlighted the predictive value of EZ layer integrity for visual outcomes at 1, 3, and 6 months post-treatment using OCT B-scans. Jee et al. [64] employed a ResNet50-based DL model with multimodal OCT data from 832 patients with CSCR, demonstrating EZ integrity (F1 score 0.8718) and retinal thickness changes (accuracy 92%) as strong visual prognosis predictors. Pfau et al. [65], using data from the PLACE trial, applied a random forest model to predict retinal sensitivity at 7–8 months post-treatment, achieving an mean absolute errors of 3.38 dB. This highlights the impacts of baseline and treatment.

Application of AI in Predicting Treatment Outcomes

Despite individual variability in the degree of SRF absorption as an assessment metric for CSCR treatment response, Fernandez-Vigo et al.’s article [66] validated the feasibility of applying DL to baseline OCT images to predict PDT efficacy in patients with chronic CSCR. However, when categorizing efficacy into complete, partial, and no absorption, the model’s accuracy was only 53%.

In contrast, Yoo et al. [67] proposed a two-stage DeepPDT-Net, in which they combined transfer learning with multimodal data (fundus photograph depth features and OCT/clinical parameters) to predict 1-year PDT outcomes (sensitive vs. resistant). The integrated model achieved an accuracy of 88%, significantly outperforming models based solely on imaging. Feature importance analysis indicated that fundus photograph depth features contributed most, followed by central foveal thickness and age, highlighting the value of multimodal fusion and transfer learning in enhancing predictive accuracy.

Beyond imaging features, research also used quantitative OCT parameters with ML to predict treatment outcomes. In a prospective 3-year follow-up study by Arrigo et al. [68], baseline OCT was used to extract metrics such as retinal thickness, RPE thickness, choroidal thickness, Sattler layer thickness, and the count of HF points at the retina–choroid interface, with key thresholds identified via decision trees. Results showed that HF count and age were the strongest predictors. The article demonstrated that combining HF quantity with patient age as stratification criteria could identify optimal treatment strategies across different age groups, providing clear and interpretable quantitative guidance for clinical decision-making.

Furthermore, recent research has applied GANs to prognosis prediction in CSCR [69]. For instance, GAN-based models generate predicted OCT images at 3-month follow-ups from baseline scans, visually simulating future lesion anatomical changes. This approach provides clinicians with an intuitive reference for prognosis prediction. The recurrence and prognosis prediction sections of Table 2 contain further details.

Research on Practical Application of AI Platforms

Initially, Kim et al. [70] used ultra-widefield indocyanine green angiography (UWF ICGA) images to classify choroidal thickening diseases with 89.19% accuracy, aiding CSCR diagnosis by providing clear lesion visualization. Subsequently, Desideri et al. analyzed OCT features of 144 patients with CSCR and identified that HF and flat irregular pigment epithelial detachment (FIPED) as biomarkers correlated with disease progression [71, 72]. In a follow-up study, their AI platform detected PED features and helped distinguish acute from cCSCR [73]. The AI platform application section of Table 2 contains further details.

Discussion

Limitations of AI in CSCR with Mitigation Strategies

Though AI models can generate ophthalmological diagnoses rapidly and accurately, integrating AI with CSCR applications faces challenges. CSCR lacks a standardized classification system, causing research and clinical practice heterogeneity. Urgent interdisciplinary collaboration is needed to integrate diagnostic and therapeutic guidelines.

Current CSCR AI models demonstrate constrained generalizability due to single-source data dependency, binary outcome limitations, and interobserver and intraobserver variability stemming from non-expert dataset annotations. These limitations compromise clinician trust in complex differential diagnoses, further exacerbated by static interpretability approaches that fail to capture real-time lesion dynamics. Future research should focus on developing multicenter, multimodal architectures integrated with dynamic visualization frameworks, which would thereby improve both interpretability and diagnostic precision in evolving pathologies.

Clinical Translation Challenges and Integration Strategies

Integrating AI models into clinical practice for CSCR presents several challenges, necessitating both technical adaptation and systemic transformation [74]. Although frameworks such as Fast Healthcare Interoperability Resources (FHIR) are theoretically ideal, the heterogeneous infrastructure across hospitals complicates the widespread adoption and compatibility of AI tools [75]. Significant variations in PACS versions, structures, and imaging device interfaces across different healthcare institutions further exacerbate integration difficulties [76, 77]. Recent advancements in intelligent adaptive systems, utilizing dynamic mapping technology, have been shown to reduce deployment time by up to 50% [78]. However, integrating data from multiple heterogeneous sources—such as the multimodal fusion of OCT, FFA, and ICGA—remains complex, and more efficient solutions are still needed to address these integration challenges [79].

Decision Support for Physicians and Patients is Central to AI Clinical Implementation

AI should serve not only as an auxiliary diagnostic tool but also provide decision support to enable more precise and personalized therapeutic decisions [80]. Despite this potential, physician trust remains a significant barrier, with concerns that AI may disrupt traditional clinical workflows and compromise care quality [81, 82]. AI should not replace physicians; rather, it should augment their capabilities. By automating routine tasks through continuous human–AI dialogue, clinicians can focus on more complex issues, improving decision-making accuracy [77, 83, 84]. Crucially, AI applications in managing CSCR must demonstrate measurable improvements in patient outcomes. This requires extending beyond diagnostic efficiency to support treatment selection, treatment response assessment, and recurrence prediction. Adherence to CONSORT-AI/SPIRIT-AI guidelines is essential for ensuring clinical evidence standards in randomized controlled trials (RCTs) and prospective studies that assess clinical feasibility and reliability [67, 85].

Regarding physician and patient acceptance, physicians are primarily concerned about the “black box” nature of AI models, particularly in the differential diagnosis of complex cases [86]. On the other hand, patients focus more on data privacy, with a significant proportion refusing to include their retinal images in AI training datasets. To address this, federated learning—by enabling distributed training of global models while ensuring that raw patient data never leaves local institutions—offers a privacy-preserving solution for multicenter data collaboration [87–89].

Explainable AI: Building Trust for Clinical Decision-Making

Enhancing model explainability is critical to establishing physician trust and ensuring the clinical applicability of AI. As deep learning technologies are increasingly integrated into ophthalmology practice, researchers prioritize transparent decision pathways, enabling clinicians to understand and verify AI-generated outputs [90].

Visualization techniques, such as attention-based heatmaps and gradient-weighted class activation mapping (Grad-CAM), intuitively highlight critical regions during the analysis of fundus or cross-sectional OCT images [23, 91]. This enables clinicians to verify whether the model accurately focuses on the characteristic pathological features of CSCR, including SRF accumulation and RPE layer abnormalities. Such transparency significantly enhances the clinical acceptability of interpretations derived from the model in practice.

Moreover, feature importance analysis methods, such as Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME), quantify the contribution of key clinical or imaging features (e.g., leakage point morphology, choroidal thickness variations) to model predictions [67, 90]. This interpretability enables clinicians to understand the rationale behind AI’s decisions, facilitating better evaluation of model recommendations in clinical decision-making.

Consequently, implementing explainable AI (XAI) techniques not only enhances model transparency and trustworthiness but also encourages physician adoption and collaboration with AI tools. This is essential for integrating AI into clinical practice and represents a critical step toward the effective clinical translation of AI in CSCR diagnosis and management.

Conclusion

AI has demonstrated substantial potential in CSCR diagnosis, evolving from basic single-dimensional analyses to complex multidimensional tasks. Current research predominantly focuses on disease classification (distinguishing CSCR from normal or other pathologies, acute versus chronic subtypes) and lesion segmentation, particularly fluid regions via OCT imaging. While OCT-based AI models show high accuracy, their reliance on NRD and PED features limits insights into choroidal dynamics, highlighting the need for an expanded focus on choroidal characteristics to deepen mechanistic understanding of CSCR. Future efforts should prioritize multimodal architectures that integrate OCTA and choroidal characteristics to enhance diagnostic precision. Improved computational efficiency in high-dimensional spaces and dynamic lesion monitoring are essential for clinical translation. As clinicians rely on multimodal retinal images, training AI with heterogeneous datasets annotated by multiple experts could better simulate real-world decision-making.

Therapeutic prognostication remains challenging because of the absence of standardized guidelines. Prospective clinical trials are urgently needed to validate AI models in practical settings, and interdisciplinary collaboration is key to establishing consensus-driven treatment protocols. Advancing these areas will enable personalized management and deepen mechanistic understanding of CSCR progression.

However, the clinical translation of AI models faces significant challenges, including infrastructural heterogeneity (e.g., incompatible PACS systems) hindering interoperability, and limited physician trust due to “black box” decision-making. To address these challenges, future work must prioritize: (1) Dynamic visualization frameworks to visualize lesion progression; (2) Standardized guidelines (CONSORT-AI/SPIRIT-AI) for validating clinical utility; (3) XAI techniques (e.g., Grad-CAM, SHAP) to enhance transparency and clinician acceptance; (4) Privacy-preserving federated learning for integrating medical data from multiple centers while ensuring patient confidentiality.

Author Contributions

Conceptualization: Wei Chi, Weihua Yang; Literature search: Ping Zhang, Qing Zhang; Data curation and formal analysis: Ping Zhang, Qing Zhang, Xinya Hu; Writing – original draft preparation: Ping Zhang, Qing Zhang; Writing – review and editing: Xinya Hu, Weihua Yang; Supervision and funding acquisition: Wei Chi, Weihua Yang. All authors read and approved the final manuscript.

Funding

Funding for the study, and the Rapid Service Fee, was supported by Shenzhen Science and Technology Program (JCYJ20240813152704006) and Sanming Project of Medicine in Shenzhen (SZSM202411007).

Data Availability

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.

Declarations

Conflict of Interest

Ping Zhang, Qing Zhang, Xinya Hu, Wei Chi, and Weihua Yang declare that they have no competing interests.

Ethical Approval

This article is based on previously conducted research and does not contain any new research with human participants or animals performed by any of the authors.