QUARTZ - Demo

To get started, click the 'Create New Session' tab using the already generated session ID and upload your images. A maximum of 100 images in total is allowed to be processed in a single session. Once all images are uploaded, click the 'Start Processing' tab. If the server is available, your job will begin immediately; otherwise, it will be placed in a queue. There is no need to remain on the page, simply note down your session ID to check your job status later. Once your job is completed, the results will be available for download for 30 days.

Colour retinal images with a 30° to 60° field-of-view angle are recommended. However, it is important to consistently use the same field-of-view angle throughout a study.

Please use only one dot in image filenames, placed immediately before the file extension.

Accepted image formats: JPEG, PNG, TIFF, and BMP.

The use of ‘Resize images’ is optional. In datasets where images have differing dimensions, it standardizes image dimensions and ensures the outputs are representative of a single image dimension.

If the ‘Resize images’ option is selected, both width and height dimensions must be specified. The selected resize dimensions will provide a target aspect ratio (i.e., the ratio of width to height); it is important that this matches the dataset's predominant aspect ratio. Outputs from images whose original aspect ratio deviates from the target aspect ratio will be excluded. Such deviations often result from cropping or unusual capture settings and will result in resizing causing uneven scaling.

Processing images in a large batch is recommended. This minimizes Docker and model startup overhead and speeds up processing through deep learning batch processing and multi-threading.

Outputs will be provided in a ZIP file containing the following folder structure:

├── Included/
│   ├── MASK/
│   ├── OD/
│   ├── Seg/
│   └── Qua/
│
└── Excluded/
    ├── MASK/
    ├── OD/
    ├── Seg/
    └── Qua/
  

Outputs are sorted into ‘Included’ and ‘Excluded’ folders. Outputs from images with a low image quality score (<= 0.69), original image dimensions below 500 pixels, or an original aspect ratio that deviates from the target aspect ratio (when the ‘Resize images’ option is selected) will be placed in the ‘Excluded’ folder. The image quality score is based on the quality of the vessel segmentation for use in epidemiological studies and not the optic disc segmentation.

MASK/: Single channel binary images of the field-of-view masks.

OD/: Three channel optic disc/cup segmentation maps where each channel is an independent binary segmentation map for a specific structure: the first channel corresponds to the optic disc, the second to the optic cup, and the third channel is blank (not used). However, users may prefer to infer structure from colour: red = optic disc, yellow = optic cup (also part of the optic disc), and black = background.

Seg/: Three channel arteriole/venule segmentation maps where each channel is an independent binary segmentation map for a specific structure: the first channel corresponds to arterioles, the second to venules, and the third to all vessels. However, users may prefer to infer structure from colour: magenta = arteriole, cyan = venule, white = crossing, blue = uncertain, and black = background.

Qua/: CSV files containing the quantitative data measured from vessel segments. There are two CSV files for each image. The first (with the suffix '_1.csv') contains image details and summary information on each vessel segment. The second (with the suffix '_2.csv') contains detailed information on each vessel segment. Coordinates in these files use a coordinate system that assigns the first coordinate to the row and the second to the column, with the top-left as the origin at (1,1). All length and area measurements are in pixels.

QUARTZ output measurements have been used to demonstrate associations with a range of phenotypes and to predict risk of disease. This section offers guidance for analysing QUARTZ outputs. However, users should also feel free to construct their own procedures informed by their specific vessel distributions.

This latest version of QUARTZ has an improved segmentation module, with more small vessels captured and represented in the vessel distribution. Therefore, previously reported associations may be altered.

• Pixel to micron conversion factor

  • To convert pixels measurements to microns, an accurate conversion factor can be determined using the camera magnification together with the axial length and/or refractive error of the eye [1]. Otherwise, an approximate conversion factor can be calculated by using the vertical optic disc size [2].
  • Pixel-based measurements can be used directly in association or prediction models without conversion to microns, with ‘Resize images’ enabled if the images have differing dimensions. However, if the refractive range of participants is wide (e.g., more than -6D of myopia to more than +6D of hypermetropia), without corrections for magnification (using axial length or refractive error methods), some degree of measurement error is likely to be introduced.

• The arteriole/venule (A/V) outputs (labelled ‘Prob A’ and ‘Prob V’ in the .csv file) are presented as probabilities of the vessel being an arteriole or a venule. We recommend excluding vessel segments with a probability < 0.8 to ensure high certainty of correct vessel type identification.

• Vessel diameter (width) outputs

  • To ensure consistent measures, exclude vessel segments with a mean diameter (labelled ‘Mean Diameter’ in the .csv file) > [mean + 4 standard deviations]. The mean and standard deviation used to derive this threshold should be determined using all vessel segments from your dataset, hence prior to the A/V probability threshold being applied.
  • If analyses require a focus on large retinal vessels rather than pre-and post-capillaries in the retinal image, we recommend excluding vessel segments with a mean diameter < 50 microns after pixel to micron conversion.

• Tortuosity outputs

  • Use the T2 tortuosity measure (labelled ‘T2: Level-Wise-Mean’ in the .csv file) for best performance [3].
  • Apply natural log-transformation to normalise the distribution, ln(T2*1000).
  • Omit tortuosity measures for vessel segments with NaN (Not a Number) or -Inf (negative infinity) values. These vessel segments are not fully excluded as they may have other valid vasculometry measures.

• Vessel area can be calculated for a vessel segment using its mean diameter * length (labelled as ‘Mean Diameter’ and ‘Length’ in the .csv file).

• Following the application of the above thresholds, exclude images if output data are provided on venules or arterioles only.

• If calculating image-level means, the contribution of each vessel segment should be weighted by its segment length.

• If using optic disc or cup size outputs, then further exclusion criteria should be considered.

  • Exclude measures with an optic disc vertical diameter (labelled ‘OD vDiameter’ in the .csv file) > [mean + 4 standard deviations] or an optic disc vertical diameter < [mean - 4 standard deviations]. The mean and standard deviation used to derive this threshold should be determined using all optic disc measures from your dataset. If required, a similar rule can be applied to the optic cup diameter.
  • If the vertical optic disc measurements are being used for pixel to micron conversion, the above exclusion criteria should be applied to exclude the image, as the conversion factor would be unreliable. Otherwise, only the optic disc and cup size outputs should be excluded.

[1] A. G. Bennett, A. R. Rudnicka, and D. F. Edgar, ‘Improvements on Littmann’s method of determining the size of retinal features by fundus photography’, Graefe’s Archive for Clinical and Experimental Ophthalmology, vol. 232, no. 6, pp. 361–367, 1994.

[2] H. A. Quigley, A. E. Brown, J. D. Morrison, and S. M. Drance, ‘The size and shape of the optic disc in normal human eyes’, Archives of ophthalmology, vol. 108, no. 1, pp. 51–57, 1990.

[3] C. G. Owen et al., ‘Measuring retinal vessel tortuosity in 10-year-old children: validation of the computer-assisted image analysis of the retina (CAIAR) program’, Invest Ophthalmol Vis Sci, vol. 50, no. 5, pp. 2004–2010, 2009.

The application is under active development, with new features (including fractal dimensions, branching angle analysis, and an optic disc segmentation quality score) planned for release in late 2025.

This web tool is supported by basic computational hardware. For high-throughput or large-scale retinal dataset processing (e.g. more than 10,000 images), please contact the team directly.

The web interface is optimized for easy access. A standalone version of the software with expanded functionality provides more detailed visualizations and advanced control. This version is not generally available.

For inquiries not covered in the documentation or citation below, such as collaboration proposals or specialized support, please contact the team at miqa@kingston.ac.uk.

The publication supporting this latest version of QUARTZ is currently under submission. In the meantime, please cite the following publication when using this tool:

R. A. Welikala, J. Fajtl, G. Johnson, F. Rahman, R. Podoleanu, P. Remagnino, E. E. Freeman, R. Chambers, L. Bolter, J. Anderson, A. Olvera-Barrios, A. Warwick, P. J. Foster, R. Shakespeare, R. Ganguly, C. Egan, A. Tufail, C. G. Owen, A. R. Rudnicka, and S. A. Barman, ‘Artificial intelligence-enabled retinal vasculometry at scale utilizing the UK Biobank, CLSA, and NEL DESP datasets’, in 2024 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI), IEEE, 2024, pp. 1–8.

Full text available via the Kingston University Research Repository

The team wish to acknowledge the funding contributions to the development of this research:

Wellcome Trust (224390/Z/21/Z), Canadian Institutes of Health Research (480750), British Heart Foundation (PG/15/101/31889), Medical Research Council (MR/L02005X/1), BUPA (F/33a/05), University of Copenhagen, City St George’s University of London, and Kingston University.