Abstract
Purpose :
Multispectral Imaging (MSI) allows for an examination of individual anatomic components of the retina via multiple monochromatic LED-sourced wavelengths during imaging. In clinical practice, ophthalmologists interpret MSI images by analyzing qualitatively the spectral consistency of MSI, i.e. the image variation across spectra. We aim at developing a quantitative tool for measuring MSI spectral inconsistency without any supervision, which is validated to be able to indicate tissue's pathological property.
Methods :
54 sequential slices were captured by leveraging an Annidis RHATM imaging device. Each sequence contains 11 spectral slices from discrete monochromatic light sources. They are binocular images from 22 patients with diabetic retinopathy and 5 healthy subjects. Their file format is dicom with a 16 bit depth and the resolution is 2048x2048. We proposed a novel unsupervised technique for the detection of retinal lesions within a set of MSI images by mathematically defining spectral consistency as pixel-specific latent feature vectors and a spectrum-specific projection matrix and defining spectral inconsistency as the number of latent feature vectors required to reconstruct the representative features. It is founded on a probabilistic Gaussian mixture model and designed to find a maximum a posterior estimate of the projection matrix and the assignment to the latent feature vectors via a stochastic expectation-maximization algorithm.
Results :
The proposed technique outperformed the human's manual works. The average area under the curve (AUC) of discriminating the degenerated and normal pixels obtained by our technique was 0.77 while the average AUC of the ophthalmologists was 0.70. An example of the MSI slices and the segmentation produced by the proposed approach are shown in Figure 1.
Conclusions :
The quantitative measurement of spectral inconsistency is not only an indicator of the retinal degeneration but also can boost the performance of retinal lesion detection and segmentation algorithms. It is an invaluable tool for various multispectral image analysis tasks.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.