A novel AI‐powered IR‐Visible dual camera system for measuring and tracking ocular surface temperature

• Published in: Acta ophthalmologica (Oxford, England) Vol. 102; no. S279
• Main Authors: Bidaki, Ehsan Zare, Wong, Alexander, Shahsavari, Navid, Murphy, Paul
• Format: Journal Article
• Language: English
• Published: Malden Wiley Subscription Services, Inc 01.01.2024
• Subjects: Algorithms; Artificial intelligence; Cameras; Control algorithms; Cornea; Deep learning; Eyelid; Image processing; Registration
• ISSN: 1755-375X; 1755-3768
• DOI: 10.1111/aos.16017

Aims/Purpose: To report the development of a novel artificial intelligence (AI) powered, dual camera (infrared (IR)/visible) system capable of measuring and tracking ocular surface temperature (OST) over any time period. Methods: The system consists of an IR camera (Teledyne FLIR IR A655sc) and a visible (V) camera (FLIR BFS 51S5C‐C), co‐mounted on a slit‐lamp for simultaneous and overlapping fields of view (FOV); designed control algorithms; computer hardware and connecting cables. Novel algorithms are leveraged to synchronize IR and V video streams of exposed ocular surface and adnexa for image and video registration. Localisation of the cornea in V video stream obtained by deep learning (DL) AI algorithms designed for semantic segmentation of the cornea. Coordinates of V video stream segmented cornea are then used to extract corresponding OST data from the IR video stream in each video stream frame. Data analysis algorithms are used to extract OST data from IR video stream. Further image segmentation algorithms isolated specific areas of interest (AOI) for OST analysis. Results: The DL algorithm was trained using V images captured using the V camera. Image registration and segmentation errors were calculated to determine accuracy of system. Mean square error for registration was 5.03 ± 1.82 (c.0.45 mm). Mean Intersection over Union (IoU) was 97.6%, representing accuracy in identifying corneal and scleral pixels in tracked eye segmentation. Analysis software extracted rate of OST change and relative OST change compared to baseline across cornea and selected AOI. Conclusions: A novel AI‐powered system for measuring and tracking OST over time was developed. The system synchronously records IR and V video streams of the eye surface and automatically extracts OST over time. The system can track eye movements and remove artefact eyelid blink frames from the data. A consistent AOI, i.e., pupil, whole cornea, inferior half, superior half, or selected corneal region, can be selected for OST extraction and analysis over time. Experimental results show that the system can track and analyse OST change over time.