A new novel method for assessing intracranial pressure using non-invasive fundus images: a pilot study

Abstract

Arteriole and venule diameter ratio (A/V-ratio) can be measured using fundus photography. In this pilot study, we correlated changes in the intracranial pressure with the diameter of vessels of the retina. We investigated whether increased intracranial pressure (ICP) was reflected in a measurable and quantifiable distention of the venule diameter, leading to a decreased A/V-ratio. This was demonstrated by assessment of the A/V-ratio in patients already undergoing conventional ICP monitoring with a cerebral intraparenchymal pressure monitor. Our method shows a correlation between A/V ratio and ICP and suggests an easily obtainable and usable point-of-care (POC), non-invasive method to estimate the intracranial pressure without the necessity of mydriatic drugs. Furthermore, the sensitivity/specificity analysis with a cut-off of < 0.8015 A/V-ratio, showed a sensitivity of 94% [85-98%] and a specificity of 50% [34-66%] with a positive likelihood ratio of 9.0. This means that in a clinical setting there is a 94% chance of correctly identifying individuals with ICP ≥ 20 mmHg.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Experimental Setup and image. (A) The experimental setup. The EpiCam M is a hand-held digital retinal fundus camera, which is aimed directly at the pupil, from where the vessels can be visualized. This figure shows a schematic setup of the experimental setup. The video capturing occurs during infusion test through a needle in the lumbar region. (B) The optic disc and vessels. To distinguish arterioles and venules it is important to know two things: 1. Arterioles never cross arterioles and venules never cross venules and 2. Venules are generally larger and darker than arterioles. If the images are of high enough quality one can observe pulsation through the arteriole and thereby making proper vessel identification easy.

Figure 2

Video processing and software analysis. (A) ‘Is Fundus’ function determines whether the image is that of a fundus (true) or not (false). This step excludes all non-fundus images. (B) The optic disc is then localized in order to measure the vessels from the same distance from the center on all images (2 radiuses). (C) Quality parameters filter images based on image quality (scores the image between 0–100). (D) Rotation: The image with the highest quality was chosen as reference, Max(Q). All images are perceived as a coordinate system (x, y, α), where α was the rotation All images were rotated to match the reference image to achieve more accurate measurements of the same points. (E) Twenty-five points are chosen for each vessel and the diameter in pixels is measured. (F) An image of how the CNTK deep learning algorithm recognizes vessels of sufficient quality on 64 × 64 px images for analysis.

Figure 3

Results: Fitted data for each patient. (A) Graphical display of the mixed-effect linear regression of Low ICP (< 15 mmHg) based on 40 observations from 11 patients with an average of 3.6 observations per patient. Each patient is represented with a fitted line if they have two or more observations. (B) Same as A. except the model is based on ICP ≥ 15 mmHg (medium ICP 15–19 mmHg and high ICP ≥ 20 mmHg. Based on 46 observations from 9 patients with an average of 5.1 observations per patient. Each patient is represented with a fitted line if they have two or more observations. Some patients have observations on both sides of the ICP spectrum (Low and Medium–High).