Physics based image interpretation to aid the detection of early signs of retinopathies

EPSRC Grant GR/S09906/01

The team

Antonio Calcagni, Ela Claridge, Jon Gibson, Mark O'Dwyer, Felipe Orihuela Espina, Iain Styles
School of Computer Science, The Unversity of Birmingham and Department of Ophthalmology, Heartlands Hospital, Birmingham

Overview

The pupil of the eye provides an opening through which the interior of the eye (ocular fundus) can be examined. The colours observed in images of the fundus depend on the architecture of its layers and the optical properties and quantities of the pigments, such as haemoglobins (in blood) and melanin (the same pigment which protects the skin from UV radiation). Many abnormal conditions cause changes in the fundus colouration, but are sometimes difficult to detect by visual inspection.

The fundus colours can be predicted from the parameters describing the quantities of the pigments residing in different layers using a physics-based simulation of the light interaction with the ocular tissue. We have established that every combination of different amounts of the main pigments gives rise to a characteristic and unique colour. This has allowed us to develop a computational method for inferring the quantities of the pigments from fundus images taken over the range of wavelengths in the visible light spectrum. The estimated quantity of each parameter in each layer is shown in a separate image called parametric map.

The maps of Retinal Blood and Macular Pigment are of particular importance in the diagnosis of eye diseases. Small haemorrhages in the retina (where cells converting light to nervous signals reside) can be initial signs of diabetic retinopathy where early detection is clinically difficult but crucial if blindness is to be avoided. The changes in the distribution of the Macular Pigment may be a predictive factor in the risk of developing sight-threatening complications in age-related macular degeneration (ARMD).

Our method for the derivation of quantitative parameters using physics-based modelling is generic and has already been applied in other imaging domains, including cancers of the skin and the colon, fluorescence microscopy and astronomy.

Key advances

The aim of the project was to develop and validate a new physics-based image interpretation method for the ocular fundus and to evaluate its clinical potential. The project has developed a detailed optical model of fundus reflectance under diffuse illumination using Monte Carlo simulation to solve the radiative transport equation. The structural layers of our fundus model include the lens, the ocular media, the neural retina, the Retinal Pigment Epithelium (RPE), the choroid and the sclera. Each layer is characterised by its thickness, refractive index, anisotropy factor, absorption coefficient and scattering coefficient. The optical properties of each layer are determined by its constituents. The underlying properties of the tissue can be regarded as constant, with the main variability being due to changes in the concentration of haemoglobins in the retina and the choroid; melanin in the RPE and the choroid, and macular pigment in the neural retina. As a refinement over the existing models, we have shown that our model can separate the contributions from retinal and choroidal blood, and choroidal and RPE melanin.

Variation of the remitted spectra with respect to changes in the concentration of retinal pigments

Macular Pigment Retinal Haemoglobins RPE melanin
Choroidal Melanin Choroidal Haemoglobins

A crucial advance has been the development of a method for the recovery of parameters (a "model inversion") characterising the ocular fundus without the need for image calibration. The method is derived from an analysis of the imaging process and is completely generic. It has been subsequently applied to other imaging problems .

We have developed a multispectral image acquisition system which can capture images from large areas of the ocular fundus at wavelengths ranging from 400nm to 1100nm. Multispectral image data sets have been collected for healthy young volunteers from a variety of ethnic backgrounds including Caucasians, African, Afro-Carribean and Mediterranean. All technically sound image sets are available for other researchers from the Web repository

By applying the inverse model to every point in a multi-spectral image set, we can compute the values of the model parameters which describe the fundus tissue at that point and construct parametric maps which show the variation of each parameter across the entire image. Figures below show RGB images and the corresponding parametric maps for the Macular Pigment and the Retinal Haemoglobins. The maps are consistent with the anatomy and physiology of the fundus.

RGB images and their parametric maps (bright=high level)

RGB image of a normal subject. Parametric map of the Macular Pigment; note the elevated levels of the pigment in the foveal area.
RGB image of a normal subject. Parametric map of the Retinal Haemoglobins (RHb); retinal vessels can be clearly seen; note the decreased levels of RHb in the foveal avascular zone.
RGB image of a patient with retinal haemorrhage. Parametric map of the Retinal Haemoglobins. note the elevated levels of RHb in the foveal area where retinal haemorrhage was diagnosed.

Publications arising from the project

Styles IB, Calcagni A, Claridge E, Orihuela Espina F, Gibson JM (submitted) Quantitative analysis of multispectral fundus images. Medical Image Analysis . Also available as Technical Report CSR-05-08, School of Computer Science, The University of Birmingham, 2005. CSR-05-08.pdf

Calcagni A, Styles IB, Claridge E, Orihuela Espina F, Gibson JM (2005) Multispectral fundus analysis. European Association for Vision and Eye Research Meeting EVER'05, Vilamoura, Portugal, October 2005. Abstract in Ophthalmic Research 37 (S1), 72. Abstracts.pdf and Poster (pdf.gz)

Preece S, Styles I, Cotton S, Claridge E, Calcagni A (2005) Model- based parameter recovery from uncalibrated optical images. Medical Image Computing and Computer Assisted Intervention (MICCAI 2005) Palm Springs, California, October 2005. LNCS vol. 3750, 509-516. Conference paper (pdf)

Styles IB, Claridge E, Orihuela Espina F, Calcalgni A, Gibson JM (2005) Quantitative interpretation of multispectral fundus images. Medical Imaging 2005: Physiology, Function, and Structure from Medical Images, Proceedings of SPIE Vol. 5746. Amini AA, Armando Manduca A Eds. (SPIE, Bellingham, WA, 2005), 267-278. Conference paper (pdf)

Orihuela Espina F (2005) Modelling and Verification of the Diffuse Reflectance of the Ocular Fundus. PhD Thesis. School of Computer Science, The University of Birmingham.

O'Dwyer M, Claridge E (2004) Single parameter recovery from modelled spectra: A comparison of spectral filter optimisation with standard multivariate approaches. Technical Report CSR-04-10, School of Computer Science, The University of Birmingham. (to be submitted to Pattern Recognition) CS-04-10.pdf

Preece SJ, Claridge E (2004) Physics-based approach to geometry insensitive recovery of quantitative scene parameters from images. Technical Report CSR-04-08, School of Computer Science, The University of Birmingham. CS-04-08.pdf

Orihuela-Espina, Claridge E, F Styles IB (2004) Validation of a physics based model of the reflectance of the ocular fundus. ARVO, Florida, April 2004. Abstract in Investigative Ophthalmology and Visual Science 2004, 45: E-Abstract 2789. Abstract and Poster.ppt

Styles IB, Claridge E, Orihuela-Espina F (2004) Quantitative interpretation of uncalibrated fundal images. ARVO, Florida, April 2004. Abstract in Investigative Ophthalmology and Visual Science 2004, 45: E-Abstract 2792. Abstract and Poster.pdf

Orihuela F, Claridge E (2003) A new optical imaging method for the ocular fundus. Medical Imaging Research Symposium, San Diego, August 2003. Abstract in Medical Physics 30(6), 1540.

Orihuela-Espina F, Claridge E, Preece SJ (2003) Histological parametric maps of the human ocular fundus: preliminary results. Medical Image Understanding and Analysis 2003, 133-136. Conference paper (pdf)

Preece SJ, Claridge E (2004) Spectral filter optimisation for the recovery of parameters which describe human skin. IEEE Pattern Analysis and Machine Intelligence, 26(7), 913-922. pdf

Claridge E, Preece SJ (2003) An inverse method for the recovery of tissue parameters from colour images. Information Processing in Medical Imaging (IPMI), Taylor C and Noble JA (Eds.) LNCS 2732, 306-317. Springer. pdf

Articles in professional press

Roberts JP (2003) Can software help prevent blindness? Biophotonics International, January/February 2003, 18-19.

Wrage H (2002) Under the skin. Professional Engineering 15(23A): Medical Engineering Supplement, December 2002, S6-S7.

Related publications

Claridge E, Powner D, Wakelam M (2005) The analysis of fluorescence microscopy images for FRET detection. Medical Image Understanding and Analysis conference MIUA'05, Mirmehdi M (Ed), 263-270. pdf

O'Dwyer M, Claridge E, Ponman T, Raychaudhury S. (2005) Mapping the physical properties of cosmic hot gas with hyper-spectral imaging. IEEE Workshop on Applications of Computer Vision, Colorado, January 2005, 185-190. pdf

Temple R, Claridge E, O'Dwyer M, Ponman T, Raychaudhury S (2005) Mapping the physical properties of cosmic hot gas with hyper- spectral imaging. Royal Astronomical Society National Astronomy Meeting 2005, Birmingham, UK. (abstract)

Hidovic D, Claridge E (2005) Modelling and validation of spectral reflectance for the colon. Physics in Medicine and Biology 50, 1071- 1093. pdf

Database of multispectral images

As a part of the project, a set of multispectral images of the fundus of healthy normal volunteers from a variety of ethnic backgrounds has been collected.