The AROMA Dataset for Automatic Detection of Artifact Type and Severity in Retinal Optical Coherence Tomography Angiography

Abstract

Introduction

In recent years, retinal vascular imaging has attracted growing interest and is experiencing rapid technological advancements in imaging modalities such as swept-source optical coherence tomography angiography (SS OCT-A). OCT-A enables precise, noninvasive, quantitative measurements of retinal vascularization. However, it is also prone to artifacts that are challenging to detect and can significantly limit diagnostic accuracy and biomarker reliability.

Methods

We present a dataset of en face retinal SS OCT-A images, labeled by artifact type and severity. Each patient’s OCT-A scan consists of 14 images that are graded by multiple experts and assigned corresponding artifact labels with an overall quality grade. We also report results for two deep-learning binary classification models trained on the dataset: one based on image compression via principal component analysis (PCA) and the other using a six-channel tensor.

Results

The dataset comprised 281 OCT-A scans from 115 anonymized patients. The PCA-based model achieved a precision of 87% for image quality classification, while the six-channel model achieved 97%.

Conclusion

Using a curated dataset of en face SS OCT-A images with detailed artifact annotations, we trained a deep-learning model capable of high-precision image quality classification. These results demonstrate the feasibility of using task-specific artificial intelligence for quality assessment based on artifact detection in retinal imaging.

Introduction

In recent years, retinal vascular imaging has attracted growing interest given the rapid technological advancements in imaging modalities, especially swept-source optical coherence tomography angiography (SS OCT-A) [1]. The retinal vascular metrics measured by OCT-A have been reported to correlate with the detection, severity, and progression of various ocular pathologies, such as diabetic retinopathy [2, 3], vascular occlusion [4, 5], and glaucoma [6, 7]. These metrics also appear to be correlated with systemic vascular diseases [8, 9] and their prognosis (acute myocardial infarction in particular [10]) as well as with multiple sclerosis [11] and Alzheimer’s disease [12]. OCT-A has gained significant importance in the emerging field of oculomics, as it enables precise, noninvasive, quantitative measurements of retinal vascularization [13, 14]. However, it is also prone to artifacts, which can be challenging to detect and can significantly hinder accurate diagnosis. The success of OCT-A-based algorithms in ophthalmology and oculomics research is closely linked to the quality of the images included in the corresponding database.

Differences in OCT-A devices and imaging protocols pose challenges to consistent artifact detection. The lack of standardized quality criteria makes it harder to develop reliable artificial intelligence models for screening, diagnosis, and monitoring based on OCT-A imaging [15]. Despite extensive research on image correction algorithms by manufacturers and academics alike [16–22], most OCT-A scans – up to 97%, regardless of in-built artifact correction algorithms [23] – are still degraded by artifacts of varying type and severity. Motion artifacts illustrate this limitation well. Many are partially corrected using hardware solutions such as eye tracking or software approaches such as multiple scan merging [24, 25]. Even in devices incorporating proprietary correction methods, Holmen et al. [23] reported motion artifacts in 93.1% of 406 OCT-A images, with alterations severe enough to yield unreliable vascular density measurements in 16% of cases. The most effective correction approaches rely on deep-learning algorithms, which are not yet used in routine practice [16, 19, 25, 26]. These models are trained on relatively small datasets and on a limited number of devices, which restricts their generalizability and clinical adoption. Moreover, they are neither integrated into proprietary image-processing pipelines nor available as open-access tools, which reduce their uptake because of the additional time required for installation and data handling.

Quality indicators provided by machine software demonstrate good sensitivity but insufficient specificity (73% and 63%, respectively) [27]. For Zeiss devices (e.g., Cirrus^®, PLEX Elite^®), image quality is assessed using the signal strength indicator. It reflects signal reflectance, defined as the proportion of incident radiant energy reflected by a surface. The manufacturer recommends a signal strength value above 6 but does not disclose how this metric is calculated [28]. A good-quality OCT-A scan is essential for obtaining reliable qualitative and quantitative data in both routine clinical practice and academic research [29, 30].

Several authors described different types of artifacts and proposed explanations for their underlying causes [23, 31–35]. Artifacts were initially classified according to four possible causes: OCT data acquisition and generation, intrinsic ocular properties, eye movement, and OCT data processing and display [33]. A recent review of the literature categorized the artifacts as patient-related, software-related, or operator-related [31]. In this extensive review, Pasha et al. [31] detailed 15 different artifacts: motion, band, segmentation, projection, projection removal, masking, unmasking, shadow, z-offset, tilt, refraction shift, decentration, defocus, suspended scattering particles in motion, and fringe washout. More recently, Lin et al. [25] classified artifacts into four categories: light-propagation and signal intensity-related artifacts, motion artifacts, artifacts caused by suboptimal imaging conditions during the acquisition process, and artifacts caused by incorrect signal processing.

Lauermann et al. [35] created a score for segmentation and motion artifacts; subsequently, Holmen et al. [23] further classified each type of artifact according to its severity on a scale from 0 to 3. As several studies have demonstrated, artifact classification remains subjective, with notable inter-operator and intra-operator variability [27, 35]. Moreover, examining several dozen images consecutively greatly increases the risk of human error through incorrect categorization of images that contain one or more artifacts. To ensure a reliable severity score for each artifact type, a repeatable, reproducible grading technique and confidence scores are required for the development of robust future OCT-A imaging datasets.

There is sparse literature on the use of CNNs for the classification of OCT-A image quality, whether used alone or integrated with conventional machine learning techniques, with only a few studies having addressed this application to date. Laueurmann et al. [36] developed a multilayer deep CNN to detect OCT-A artifacts and classify 200 images as either good or poor quality, with 92% accuracy. Their analysis was limited to images from the superficial capillary plexus and focused only on motion and segmentation artifacts. More recently, Dhodapkar et al. [37] used a ResNet152 neural network classifier, pretrained using ImageNet, to classify images of the superficial capillary plexus in 347 OCT-A scans. They trained two models, one to identify high-quality images (maximal prediction accuracy of 95.3%) and one to identify low-quality images (maximal prediction accuracy of 93.5%). Quality was described according to a composite score, based on the visibility of small capillaries and the severity and prevalence of motion, segmentation, centering, and opacity artifacts. The purpose of this study were (a) to release a publicly available database, AROMA (artifact recognition on OCT-A, modeling, and analysis), consisting of en face retinal SS OCT-A images labeled by artifact type and severity and (b) to provide two deep-learning models designed to assess quality based on artifact detection.

Methods

Data Organization

The dataset released with this study is a new publicly available retinal SS OCT-A dataset labeled according to the severity of each type of artifact [38]. For each patient, the dataset included 14 images derived from a single OCT-A scan, the signal strength of the OCT-A, and the severity grades of seven labeled artifacts (all detailed below). Detailed information on the dataset acquisition is provided in Table 1. A sample of the 14 images captured from each patient is presented in Figure 1.

Table 1.

Dataset content of the AROMA study: acquisition modalities and data characteristics

Subject area	Biomedical imaging, ophthalmology
Type of data	Image, table
Method of data acquisition	Swept-source OCT-A: PLEX Elite 9000® (Carl Zeiss Meditec Inc., Dublin, OH, USA)
Data format	PNG, CSV
Experimental factors	Pupillary dilatation with tropicamide 0.5%
Experimental features	Macular angiography 3 × 3 mm and 6 × 6 mm, 100 Hz, signal strength ≥7
Data source location	University Hospital of Dijon, Dijon 21000, France

Fig. 1.

Sample images from the dataset. Names are derived from the proprietary ARI-Network^® algorithm. a Superficial structural enface. b Superficial Angio enface. c Superficial Angio enface vessel trace. d Superficial Angio enface vessel map overlay. e Deep structural enface. f Deep Angio enface projection removed. g Deep Angio enface projection removed vessel trace. h Deep Angio enface projection removed vessel map overlay. i Retina structural enface. j Retina Angio enface. k Retina Angio enface vessel trace. l Retina Angio enface vessel map overlay. m Retina slab with FAZ. n FAZ area map. A single OCT-A scan was decomposed in 14 images. Among them, six were selected and served as input in the binary classification models: a–c, e–g. OCT-A, optical coherence tomography angiography; FAZ, foveal avascular zone.

Data Origin

This dataset was acquired from the Department of Ophthalmology at the University Hospital of Dijon, France, and consists of acquisitions from three registered clinical studies. It was anonymized and processed in accordance with the rules established by the Ethics Committee of the University Hospital of Dijon. All administrative information contained in the metadata was removed, making the dataset untraceable and fully anonymized.

Selected patients did not have any macular disease or severe ocular media alteration. In eyes without maculopathy, the retinal vascular network remains sufficiently preserved to allow reliable extraction of OCT-A derived biomarkers. In contrast, maculopathies induce major vascular alterations that limit the relevance of these biomarkers for oculomics. Only one eye per patient was analyzed.

OCT-A Scans

All OCT-A scans were performed using the same PLEX Elite 9000^® device (Zeiss, Carl Zeiss Meditec Inc., Dublin, OH, USA) after instillation of one drop of tropicamide 0.5%. One OCT-A scan was collected per visit, with a maximum of two visits per participant. OCT-A image sizes were 3 × 3 mm and/or 6 × 6 mm. If both scan sizes were available for a participant, each scan was considered a separate instance in the analysis. All OCT-A scans were acquired with a frequency of 100 Hz. Only scans with a signal strength ≥7 were included. The PLEX Elite 9000^® acquisition sequence includes in-built artifact correction algorithms, notably for motion and projection [39, 40]. In total, 281 OCT-A scans were acquired from 115 patients as part of the AROMA dataset.

En Face and Layered Image Acquisitions

Swept-source OCT-A scans the retina to generate 2D images, called “B-scans,” which are then assembled into a 3D image. From this, en face images can be extracted, similar in appearance to a conventional retinal angiogram. Depending on the depth layer selected during data acquisition, the en face image represents a deeper or shallower slab of the retina. Subsequently, several proprietary image-processing techniques are applied to the same en face image to segment vessels or the foveal avascular zone (FAZ). In our study, only en face images were assessed, as they represent the primary source for quantitative analysis of retinal microvascularization. After acquisition, each OCT-A scan was exported to the proprietary cloud platform ARI-Network^® (Carl Zeiss Meditec Inc., Dublin, OH, USA), where it was processed using a dedicated algorithm (Macular Density v0.7.3) [41]. The algorithm computed each OCT-A scan into 14 different images according to the layer studied and the type of image-processing performed (Fig. 1). Depending on the initial acquisition size (3 × 3 mm or 6 × 6 mm), the resolution of an image was either 216 × 216 pixels or 512 × 512 pixels, respectively. The layers, or slabs, as predefined by the proprietary algorithm are: (1) superficial, from the top of internal limiting membrane to the bottom of the inner plexiform layer; (2) deep, from the top of inner plexiform layer to the bottom of outer plexiform layer; (3) retina, from the top of internal limiting membrane to the bottom of Bruch’s membrane minus 41 microns (Fig. 2).

Fig. 2.

En face images (left) corresponding to the B-scan images (right) for the “retina,” “superficial,” and “deep” layers of an OCT-A scan acquired with a PLEX Elite^® device. “Retina” layer extends from the top of internal limiting membrane to the bottom of Bruch’s membrane minus 41 microns. “Superficial” layer extends from the top of internal limiting membrane to the bottom of the inner plexiform layer. “Deep” layer extends from the top of inner plexiform layer to the bottom of outer plexiform layer. OCT-A, optical coherence tomography angiography.

There were six types of image-processing techniques: (1) structural en face, i.e., en face OCT image with no angiography; (2) Angio en face, i.e., the classic en face OCT-A image; (3) vessel trace, i.e., binarization of the Angio en face image to form a black-and-white illustration of the vessels; (4) vessel map overlay, i.e., averaging of the vessel trace image, which is then displayed as a heat map with an ETDRS grid; (5) slab with FAZ, i.e., automatized segmentation of the FAZ from the Angio en face image; (6) FAZ area map, i.e., binarized mask of the FAZ segmented on the slab with FAZ image.

The combination of these three layers and six image-processing techniques generates 14 images from one OCT-A scan. All images were inspected for artifact labelling. However, not all of them were used as model inputs. Some of these images are very similar, and therefore we selected only six to input into the model to avoid redundant information: superficial structural en face, superficial Angio en face, superficial vessel trace, deep structural en face, deep Angio en face, and deep vessel trace. These images were selected because they were those in which examiners identified the highest number of artifacts. They are also minimally altered by the proprietary ARI-Network^® image processing, unlike vessel map overlay or FAZ area map images, which undergo more extensive transformations.

Artifacts and Quality Labels

The images were reviewed from the ARI-Network^® cloud service. Using a grading system derived from Holmen et al. [23], the following artifacts were rated from 0 to 3: “Motion/Blink,” “Tilt,” “Focus,” “Decentration/Z-offset,” “Shadow/Masking,” “FAZ segmentation,” and “Refractive Shift” (Table 2). A grade of 0 indicates the absence of artifacts, while a grade of 3 denotes more severe artifacts, based on their extent (% of en face OCT-A scan lines) and central involvement. This severity scale is consistent with those used in SD-OCT evaluation guidelines [23]. Definitions of the artifacts are provided in Table 2. They were assessed independently for artifact presence and severity by two experienced ophthalmology fellows (L.S. and P.D.). In the event of a discrepancy, the final decision was made by a senior ophthalmologist (L.A.). The same OCT-A image may be affected by different artifacts. Examples of artifacts are shown in Figure 3.

Table 2.

OCT-A image artifacts and corresponding grades for each image set, adapted from Holmen et al. [23]

Artifact	Definition	Severity grade
		0	1	2	3
Movement or blink	Eye movement artifact is seen as thin white lines in conjunction with interruption, displacement, doubling or ghosting of vessels. Blink artifact appears as horizontal black bands on the angiogram.	No vertical or horizontal white lines, interruption, displacement, or doubling of vessels	Interruptions ≤1% of lines within ETDRS grid	Interruptions ≤10% of total lines or ≤10% of lines within inner subfield	Interruptions >10% of total lines or >10% of lines within inner subfield
Defocusa	Global loss of image sharpness with diminished visibility of small capillaries.	Well-defined vessels on angiogram in comparison with standard image	Decreased definition of small capillaries in the center and inner subfield without affecting overall VD map	Same with ≤0% of overall VD map affected	Same with >10% of overall VD map affected
Tilta	Unilateral vessel dropout on angiography caused by an oblique acquisition angle, head misalignment, or high myopia.	No tilt visible	Asymmetric decreased definition of small capillaries unilaterally without affecting the VD map	Same with ≤10% of overall VD map affected	Same with >10% of overall VD map affected
Decentration or z-offset	Decentration occurs when the scan is not centered. Z-offset artifact occurs when the retina is vertically displaced outside the OCT acquisition window due to improper head positioning.	No decentration; ETDRS grid centered on fovea and within angiogram	≤1% of total lines within ETDRS grid out of view	≤10% inner subfield lines out of view	>10% inner subfield lines out of view
Shadow	Localized attenuation of retinal signal intensity, usually caused by vitreous floaters or corneal opacities.	No localized, decreased intensity with associated loss of angiogram	Localized, decreased intensity with change on angiogram and ≤1% of lines within ETDRS grid	Same with ≤10% of lines within inner subfield	Same with >10% of inner subfield OCT lines
FAZ segmentationb	FAZ segmentation appears falsely enlarged or reduced.	No errors in FAZ segmentation	Error in FAZ segmentation with no other artifact explaining it, on 1 quarter-circle	Same on 2 quarter-circles	Same on ≥3 quarter-circles
Refractive shift	Change in reflective intensity between contiguous lines owing to blinking and a change in corneal refractive index on the corneal surface.	No abrupt change in reflective intensity between contiguous lines	Abrupt change in reflective intensity. Fewer than ≤1% of contiguous lines within ETDRS grid affected	Same with fewer than ≤10% of total lines within inner subfield affected	Same with more than >10% of total lines within inner subfield affected

FAZ, foveal avascular zone; OCT, optical coherence tomography; VD, vessel density.

^aVD images required to assess this artifact.

^bFAZ images required to assess this artifact.

Fig. 3.

Different types of artifacts visible on en face OCT-A images. a Movement artifact on a “Superficial Angio Enface” image. b Tilt artifact on a “Superficial Angio Enface” image. c Focus artifact on a “Deep Angio Enface Projection Removed” image. d Decentration artifact on a “Superficial Angio Enface” image. e Shadow artifact on a “Superficial Angio Enface” image. f FAZ segmentation artifact on a “Retina Slab with FAZ” image. g Refractive shift artifact on a ”Superficial Structural Enface” image. OCT-A, optical coherence tomography angiography; FAZ, foveal avascular zone.

For each OCT-A scan, a global quality grade was used, computed as follows: (1) “Good” if the highest severity grade was 0 or 1, and the sum of the grades was less than 6; (2) “Sufficient” if the highest severity grade was 2 and the sum of the grades was less than 6; (3) “Poor” if the highest severity grade was 3, or if the sum of the grades was more than 5. This composite classification reflects the general interpretability of an image. It has two main advantages. Clinically, routine practice does not require precise identification of each artifact type, but rather assessment of whether artifacts are too numerous or too severe to allow meaningful quantitative analysis. It is also beneficial for model training, as it increases the number of images available per class, rather than multiplying classes with too few images. These advantages make composite scores a common approach in studies on artifact detection [23, 37].

Deep-Learning Models

Two deep-learning models were developed, both aiming to classify an OCT-A scan as either poor quality or not-poor-quality, though using different methodological approaches. The data flow and its integration into the two different models are summarized in Figure 4. Except for the initial processing performed on the ARI-Network^® cloud platform, images were not modified, aside from resizing to 224 × 224 pixels using the PIL library. The binarization of the dataset was achieved by merging the “Good” and “Sufficient” labels into a single “Not-poor” category. The dataset was organized such that each image was labeled with the corresponding class. We used balanced class weights to reduce imbalance between the two categories using, which were calculated during training using scikit learn library (version 1.5). The dataset was randomly split into train and validation sets in an 80%–20% ratio, respectively. The models were implemented using TensorFlow 2.17 and trained for 200 epochs with a batch size of 32. Training was performed using the Adam optimizer (learning rate 1e−4) with a cosine decay scheduler on a single V100 GPU. Early stopping was guided by accuracy, with a monitoring window of 30 epochs. The performance of the models was evaluated using the following metrics: accuracy, binary cross-entropy, binary accuracy, precision, recall, and area under the ROC curve. Confusion matrices were used to represent the models’ performance.

Fig. 4.

Overview of data flow and integration into the two classification models. For each patient, the dataset included 14 images derived from a single OCT-A scan, the signal strength of the OCT-A, and the severity grades of seven labeled artifacts. Out of the 14 images, six were selected for extracting the principal components – PCA-based binary classification model (e1) – or serving as a 6-channel tensor – six-channel binary classification model (e2). Models were trained to classify the images into two classes: not-poor-quality and poor-quality. OCT-A, optical coherence tomography angiography; PCA, principal component analysis.

The first model was trained using images generated by principal component analysis (PCA) [42] applied to each set of six images. One set corresponded to one OCT-A scan, as detailed above. PCA was used to reduce dimensionality while preserving the most informative features of the dataset. The first three principal components were retained and normalized to the (0, 1) range. Each component was then reshaped into a 256 × 256 single-channel image. Finally, the resulting components were concatenated along the channel dimension to form a three-channel tensor of size 3 × 256 × 256. A binary classification convolutional network was developed based on a VGG16 backbone trained on ImageNet to enable transfer learning of the features [43], taking the 3-channel PCA images as input. A series of fully connected layers followed, incorporating batch normalization, dropout (30% rate), and ReLU activations. The final output layer used a sigmoid activation for binary classification. Binary cross-entropy served as the loss function. A graphical overview of the model is provided in Figure 5.

Fig. 5.

Convolutional neural network with VGG16 backbone for classifying principal component analysis images. Six images were selected for extracting the principal components. The three principal components were then combined to form the three channels of a single RGB image representing each OCT-A scan. A binary classification network was developed based on a VGG16 backbone trained on ImageNet to enable transfer learning of the features. A series of fully connected layers followed, incorporating batch normalization, dropout (30% rate), and ReLU activations. The final output layer used a sigmoid activation for binary classification. Training was performed using the Adam optimizer (learning rate 1e−4) with a cosine decay scheduler on a single V100 GPU. Binary cross-entropy served as the loss function. OCT-A, optical coherence tomography angiography.

A second model was trained after stacking the OCT-A images into a single n-channel tensor. Instead of extracting the principal components of the six images, we stacked them in a single tensor of size 6 × 256 × 256. The model architecture consisted of a custom CNN with three convolutional layers, each followed by a 30% dropout layer, with the last dense layer having sigmoid activation. The model was trained with a binary focal loss function, which is a variant of the binary cross-entropy loss function that focuses on misclassified examples [44]. A graphical overview of the model architecture is presented in Figure 6.

Fig. 6.

Custom convolutional neural network for classifying images stacked as six-channel tensor. Six images were selected and stacked in a single tensor of size 6 × 256 × 256. The model architecture consisted of a custom convolutional neural network with three convolutional layers, each followed by a 30% dropout layer, with the last dense layer having sigmoid activation. The model was trained with a binary focal loss function.

Results

Overall, 281 OCT-A scans were collected from 115 patients. The proportions of artifacts in our dataset are presented in Table 3, and the image characteristics (number, resolution, quality labels) are listed in Table 4.

Table 3.

Proportion of artifacts in the OCT-A scans and the training and test datasets

Artifacta	All OCT-A scans (n = 281), N (%)	Training set (n = 245), N (%)	Test set (n = 36), N (%)
Motion/blink	68 (21.35)	60 (24.49)	8 (22.22)
Tilt	63 (17.79)	50 (20.41)	13 (36.11)
Defocus	65 (20.28)	57 (23.27)	8 (22.22)
Decentration/Z-offset	39 (11.03)	31 (12.65)	8 (22.22)
Shadow/masking	68 (20.64)	58 (23.67)	10 (27.78)
FAZ segmentation	68 (23.49)	66 (26.94)	2 (5.56)
Refractive shift	10 (2.85)	8 (3.27)	2 (5.56)

OCT-A, optical coherence tomography angiography.

^aAny artifact with a grade >0 was considered present.

Table 4.

Distribution of the training and test set, proportion of 3 × 3-mm (256 × 256 px) and 6 × 6 mm-images (512 × 512 px) in each set (px indicates image dimensions in pixels)

​	Training + validation set			Test set
	3 × 3	6 × 6	total images	3 × 3	6 × 6	total images
Good	108	72	180	5	19	24
Sufficient	10	13	23	1	5	6
Poor	14	28	42	3	3	6
Total	132	113	245	9	27	36

The PCA-based binary classification model was trained for 200 epochs, with early stopping at 153 epochs. It yielded a precision, recall, and F1 score of 0.8667, 0.8966, and 0.8814, respectively (Table 5). The true-positive rate (sensitivity) matched the recall at 89.66%; however, the true-negative rate (specificity) was 42.86%.

Table 5.

Results obtained using both binary classification models on the test dataset (bold indicates the better performance)

​	PCA-based	Six-channel
Sensitivity/recall	0.8966	0.8788
Specificity	0.4286	0.6667
Accuracy	0.8056	0.8611
Balanced accuracy	0.6626	0.7727
Precision	0.8667	0.9667
False-positive rate (type I error)	0.5714	0.3333
False-negative rate (type II error)	0.1034	0.1212
Negative predictive value	0.5000	0.3333

PCA, principal component analysis.

The six-channel binary classification model was trained for 80 epochs with early stopping criteria. It demonstrated an overall precision of 0.9667, a recall of 0.8788, and an F1 score of 0.9206 (Table 5). Sensitivity and specificity were 87.88% and 66.67%, respectively. The confusion matrices are presented in Figure 7.

Fig. 7.

Confusion matrix for PCA-based (a) and six-channel (b) binary classification models on the test dataset. The PCA-based model yielded a precision of 0.8667, an F1 score of 0.8814, a sensitivity of 0.8966, and a specificity of 0.4286. The six-channel model demonstrated a precision of 0.9667, an F1 score of 0.9206, a sensitivity of 0.8788, and a specificity of 0.6667. PCA, principal component analysis.

Discussion

This study presents a novel deep-learning approach to OCT-A quality assessment based on artifact detection, relevant for both clinical and research settings. Beyond the modeling contribution, a major novelty lies in the release of a curated OCT-A dataset, annotated by experts for artifact type and severity across multiple en face OCT-A scans. By including a wide range of artifact severities, including less frequent but clinically meaningful artifacts, it provides a foundation for future work on quality assessment methods and oculomics-oriented biomarker research.

Distribution of the Artifacts

After analyzing 406 OCT-A scans, Holmen et al. [23] found a very high prevalence of artifacts: 93.1% motion, 74.9% defocus, 62.3% shadow, 50.5% tilt, 43.8% Z-offset, 31.8% refraction shift, and 21.4% decentration. In our cohort, the corresponding rates were substantially lower (e.g., 21% motion, 20% defocus, 20% shadow, 18% tilt, 3% refraction shift, and 11% combined Z-offset and decentration). These differences likely reflect distinct study populations and methodologies. Holmen’s work included only patients with diabetic retinopathy, a group known to have more frequent and severe OCT-A artifacts due to comorbid retinal changes and fixation challenges [45]. The type of images annotated also differs: Holmen et al. [23] labeled both B-scans and en face images, whereas we annotated only en face scans, which may underrepresent artifacts best seen on cross-sections (e.g., refraction shift). Device differences may contribute as well, since Holmen used older SD-OCT-A platforms, while we used a more recent swept-source device with improved built-in correction. Finally, human grading variation may influence artifact detection, and Holmen et al. [23] did not specify pupillary dilation status, a factor known to affect image quality and artifact prevalence [30].

Model Comparison

The PCA-based model demonstrated good overall performance, with high precision, recall, and F1 score. Its high sensitivity and low false-negative rate indicated a strong ability to detect poor-quality images. However, the limited specificity suggests difficulty in correctly identifying not-poor quality images.

The six-channel model achieved higher precision and a higher F1 score. It offered a better balance between sensitivity and specificity, despite class imbalance. By preserving raw spatial information, it outperformed the PCA-based model, which may lose relevant features during compression. This model establishes a baseline and confirms the dataset’s value for future research in artifact detection and image quality classification.

To enable the use of a pretrained ImageNet model requiring a three-channel input, PCA was applied to reduce the dimensionality of the OCT-A scan inputs while preserving the most informative features. However, this compression may have removed subtle spatial information, especially for complex or infrequent artifacts. Because PCA is unsupervised, it does not prioritize features of clinical relevance. This limitation may explain why the six-channel model achieved better results. Despite its higher input dimensionality, the six-channel architecture may have captured artifact heterogeneity more effectively, which could explain its higher accuracy and more robust classification.

Adapting the Grading System

The artifact grading system proposed by Holmen et al. [23] was followed because it focuses on 10 types of artifacts and proposes a homogeneous grade from 0 to 3 for each one. Holmen et al. [23] describe it as the first available scoring system for OCT-A scans to incorporate multiple artifacts, visualized from B-scans and en face images. For our AROMA dataset, we chose to adapt their grading system (Table 2). Because we did not analyze B-scans, certain criteria were adapted for en face image analysis. Artifact categories with similar appearance and origin – such as decentration and z-offset – were merged, which also helped reduce class imbalance. Projection and segmentation artifacts were not visible on en face images. They were therefore not labeled, and no corresponding columns were included in the dataset.

Comparison with Previously Developed Algorithms for OCT-A Artifact Detection

As described above, studies by Laueurmann et al. [36] and Dhodapkar et al. [37] also aimed to develop algorithms to classify OCT-A scan quality based on the prevalence and/or severity of artifacts. However, their results are not fully comparable with ours. First, their models analyzed only images from the superficial capillary plexus, whereas we included all OCT-A layers to better reflect current research practice, as deeper layers may provide relevant information [9]. These deeper layers are more frequently affected by artifacts, a setting in which our algorithm may show improved performance. In addition, Laueurmann et al. [36] based their classification on both en face images and B-scans, while our approach relied exclusively on en face scans. Second, their models focused mainly on motion and segmentation artifacts, with additional centering and opacity artifacts in the model by Dhodapkar et al. [37] In contrast, we aimed to detect less frequent artifacts such as tilt and defocus, which are more challenging to identify and make model training more complex. The defocus artifact mainly affects the deep layers, while the tilt artifact, by definition, renders the differences between the quadrants of the ETDRS grid uninterpretable. The latter artifacts, which are often overlooked, pose a major risk of introducing bias, particularly in clinical studies. Third, the methods of data collection – different machines (Angiovue^® versus PLEX Elite^®), proprietary algorithms – and label annotation differed from the aforementioned studies.

Vision-language models (VLMs) are another type of artificial intelligence that combine visual and textual inputs to perform tasks, such as image captioning and visual question answering. In a recent study, VLMs did not reach human-level performance when detecting artifacts on color fundus photographs [46]. It would be interesting to assess the performance of VLMs for artifact detection on en face OCT-A images. Their ability to identify subtle imaging artifacts in medical datasets remains uncertain.

Strengths and Limitations

The strengths of our study included the creation of a publicly available, expert-annotated dataset of en face SS OCT-A scans with detailed artifact grading. The grading method was standardized, reproducible, and adapted specifically for en face images. Our algorithm was trained on data derived from real-world clinical studies, covering both 3 × 3 mm and 6 × 6 mm scans. Unlike prior studies, we included deep retinal layers and less frequent but impactful artifacts such as tilt and defocus, which are often overlooked yet are critical for reliable quantitative analysis.

We also acknowledge certain limitations. First, the marked imbalance in the dataset likely accounts for the low specificity (0.67). Class imbalance can lead to biased model performance, favoring majority classes and thus compromising the ability to generalize results. Second, extrapolation of the results of this study is also limited by the fact that analyses were conducted exclusively on data from the PLEX Elite^® device. Third, the ARI-Network^® algorithm used for the initial stage is proprietary, which restricts the reproducibility of results on other devices and platforms. Fourth, some types of artifacts were underrepresented, especially refractive shift, which might affect the model’s sensitivity to these cases. Finally, the categories “not-poor” and “poor” were defined arbitrarily and were not directly linked to measured biomarker degradation. As a result, biomarkers derived from scans labeled as not-poor quality may still be unreliable, and conversely. This classification reflects only the global artifact burden and does not account for the impact of specific artifacts on biomarker extraction.

Opportunities for Further Development

Several promising research directions can be pursued to enhance model performance and applicability. First, we could design a model specifically for multi-artifact detection rather than global quality classification. Second, creating an ensemble of specialized models for different image dimensions, such as 3 × 3 mm and 6 × 6 mm scans, may leverage scale-specific features and lead to improved classification accuracy. Third, we could include data from other OCT-A devices to evaluate cross-platform robustness. Fourth, the inclusion of pathological OCT-A scans in the dataset (e.g., of glaucoma, age-related macular degeneration, diabetic retinopathy) would be of substantial clinical interest [27]. Fifth, expanding the dataset with a more diverse and balanced set of images will likely enhance the robustness and generalizability of future models.

Conclusions

This study introduced a novel dataset of en face retinal OCT-A scans annotated with artifact types and their corresponding severity levels. Seven artifact types were assessed and graded on 281 OCT-A scans from 115 patients. The most frequent artifacts were “Motion/Blink,” “Shadow/Masking,” and “Defocus.” We trained a deep-learning model capable of classifying image quality with a precision of 97%. While these results are promising, further optimization is needed. External validation is essential to confirm the model’s robustness and generalizability across different populations and acquisition settings. This work lays the foundation for future studies exploring the integration of automated quality assessment tools in ophthalmic imaging workflows.

Acknowledgments

The authors thank Laure-Anne Steinberg for cleaning and organizing the clinical study data.

Statement of Ethics

The dataset was collected at Dijon University Hospital and derived from three previously registered clinical studies. All administrative information included in the metadata was removed, making the dataset untraceable. In accordance with French regulations governing anonymized retrospective data, additional ethical approval was not required. The three original studies were registered on ClinicalTrials.gov under the identifiers NCT05315349, NCT04855123, and NCT03979001. They were performed in accordance with the Declaration of Helsinki and approved by the official regional Ethics Committee “Committee for the Protection of Persons Est-III”, with approval dates of June 7, 2022, March 18, 2021, and January 6, 2020, respectively. Participant registration took place between January 2020 and April 2025 across the studies. All adult participants provided written informed consent prior to participation.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This study was not supported by any sponsor or funder.

Author Contributions

Conceptualization: F.M. and L.A.; methodology: F.M.; software: L.S., A.A., and A.G.; formal analysis: A.A.; resources: P.G., C.C., and F.M.; data curation: L.S., A.A., and P.D.; writing – original draft preparation and project administration: L.S.; writing – review and editing: L.S., A.A., F.M., and L.A.; visualization: A.G.; supervision: L.A. All authors have read and agreed to the published version of the manuscript.

Funding Statement

This study was not supported by any sponsor or funder.

Data Availability Statement

The data that support the findings of this study are openly available in Zenodo at https://zenodo.org/records/18258095 [38].