Validating the Diagnostic Accuracy of Mediview 2.0 Software as an Image Analysis Tool of Meibomian Glands

Meibomian gland dysfunction (MGD) is a chronic abnormality of the meibomian glands and commonly characterized by terminal meibomian duct obstruction and/or qualitative or quantitative changes in the glandular secretion. This may result in alteration of the composition and functioning of the tear film lipid layer, and subsequently destabilisation of the tear film. Studies have shown that MGD is one of the most common causes of evaporative dry eye.

It is recommended that the diagnosis of MGD be made by assessing ocular symptoms, lid morphology, mebomian gland mass, gland expressibility, lipid layer thickness and meibography. In meibography, the structure of the meibomian glands, including the ducts and acini can be observed. Photographic documentation of the meibomian gland is also possible under specialised illumination techniques.

There are two types of meibography currently in use- contact meibography and non-contact meibography. Non-contact meibography is advantageous over contact meibography as it is more comfortable for patients. In non-contact meibography, the slit-lamp microscope is equipped with an IR charge-coupled (CCD) device video camera and an IR transmitting filter. During the procedure, IR light is projected onto the inverted eyelid of the subject and the image of the eyelid captured by a camera.

Currently, skilled clinicians analyse images obtained via non-contact meibography manually to determine the health of the patient's eyelid. Through a collaborative project between scientists from bioinformatics institute, A*STAR, Singapore (Dr Lee Hwee Kuan) and SERI, semi-automatic algorithms for classifying meibographs into healthy and unhealthy have been previously developed. The commercial partner for this project, Mediworks, China, has incorporated these algorithms into the Mediview 2.0 software. This software consists of modules for acquiring meibography images as well as for analysing these images. Provided a user draws the outline of the tarsal plate on the captured image, the software is able to classify images into two categories- Healthy and Unhealthy without the aid of a skilled clinician. A trained technician could be taught how to capture the images and hence leaving the doctor with more time to focus on clinical assessment and treatment.

However, the accuracy and reproducibility of this software should be validated against a clinical assessor just as in the original algorithms. We propose to recruit 50 healthy volunteers and 50 patients from SNEC clinics and from public.

Clinical importance Should this modality of imaging be shown to be repeatable, it can be incorporated into current protocol/workflow for assessment and monitoring of Meibomian Gland Dysfunction and treatment progress in our centre. The images captured in this project can also be used for validating further updates of the Mediworks software, since the hardware will remain unchanged in the future.

Study Objectives and Purpose The primary purpose is to validate Mediview 2.0 software as an objective tool in assessing the health of meibomian glands.

The specific aims include assessing a., the accuracy of Mediview 2.0 in classifying meibograph images (into Healthy and Unhealthy categories) and b., reproducibility of the classifications.

Study design:

Prospective study

Rationale:

Mediview 2.0 is a newly developed software and has yet been validated as a tool for assessment of the health of meibomian glands.

Methods:

Participants and target sample size 50 healthy volunteers and 50 patients aged 21 to 99 years will be recruited for this study. They are outpatients and accompanying relatives from the Singapore National Eye Centre. Permission would be sought from the attending doctors before subjects are being recruited. Informed written consent will be obtained from all participants.

Visit schedules One visit is required for participants. In general, they will not have to come for additional visits if they are recruited whilst on a scheduled clinical appointment for other reasons. In exceptional cases, if the meibography cannot be performed on the same day, participants would be given the option of coming on another suitable day for this study.

Duration of study:

12 months.

Procedures:

1. Imaging of Meibomian Gland During the imaging of each eye, participants will be required to place their chins on a chin rest on the slit-lamp microscope. The upper eyelids of the participant will be in turn everted and an image of meibomian glands present on the eyelid will be captured. This is also performed for the lower eyelids. This is a non-invasive technique and the camera and IR lamp will not touch the patient. During each eversion of eyelid, optometrist will acquire 3 images at one time. Thereafter, Mediview 2.0 software will proceed to analyse the image into healthy and unhealthy glands. Throughout all measurements, all parameters of the meibographer, (such as light intensity, magnification of eyelid) will be kept constant.
2. Analysis of Image of Meibomian Gland The images taken above will be exported and assessed by another optometrist who is unaware of the classification of the image by Mediview. This assessor will classify the images into healthy and unhealthy. Results between the Mediview 2.0 software and the optometrist will be assessed. Should there be any disagreement on the results, a 3rd clinician will be involved in the validation.

Statistical analysis:

1. Comparison between software classification and human classification Cohen's kappa coefficient will be calculated to assess the agreement between these two ratings. For interpretation of the calculated statistic, we use recommendations from Landis and Koch, who characterized values < 0 as indicating no agreement and 0-0.20 as slight, 0.21-0.40 as fair, 0.41-0.60 as moderate, 0.61-0.80 as substantial, and 0.81-1 as almost perfect agreement.
2. Comparison between the first and second software classifications Similarly we will use the kappa coefficient to evaluate the agreements between the first and second classifications made by the software.

Expected outcomes:

The classification of meibograph images from Mediview 2.0 should show good agreement with the assessment by the skilled clinician. Similarly classifications of the meibographs taken on the same day for the same eyes should show good agreement.

Potential problems No potential problems are expected for this study.