A System for Predicting White Matter Lesion Volume in the Brain Using Retinal Vessel Features

Abstract

Purpose : To assess the accuracy of a novel computer based system for predicting the brain white matter lesion (WML) using retinal vascular features which is highly associated with stroke and dementia.

Methods : 111 patients’ data were taken from the ENVISion study (A Neuro-Vascular Imaging Study, Australia; persons aged ≥ 70 years) which has retinal and MRI images. The vessel features such as focal arteriolar narrowing (FAN) and artery vein nicking (AVN) were quantified through Canny edge detection and Graph optimization algorithms. The Dijkstra’s shortest path algorithm was applied to optimize the graph search obtaining the vessel edges to compute widths. The vessel widths were analysed to quantify the severity of AVN and FAN. WML volume was computed using pixel contrast and spatial prior knowledge, Random Forest classifier and Markov Random Field algorithms. WML volume was categorized as mild and severe and used as output classes in the model. Finally, a Support Vector Regression model was trained by using these quantified AVN and FAN to predict the volume of WML. We evaluate our proposed model on the dataset through estimating the WML volume with MSE between our predicted WML load and manually annotated WML load. For accuracy test on the classification of patients having mild and severe WML load, we compute the sensitivity, specificity and F1 score.

Results : The MSE between our predicted and manually annotated WML load is 0.13. The proposed prediction model obtains sensitivity of 0.83, specificity of 0.60, positive predictive value of 0.69, negative predictive value of 0.76, and F1 score of 0.77 in classifying the patients having mild and severe WML load.

Conclusions : The results indicate that our retinal image based WML prediction model can be helpful for the physicians to identify the patients who need immediate MRI screening for the further diagnosis of WML and regular monitoring for Stroke and Dementia.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

Figure shows an example of the segmentation outcome of the proposed white matter lesion segmentation method such as: (a) FLAIR MRI, (b) initial segmentation of lesion, (c) refined segmentation of lesion after using Markov Random Field (MRF) and (d) manual segmentation of lesion.

 

Figure shows the automatically mapped boundary of arteriolar segments having different severity levels of focal arteriolar narrowing (FAN) such as (a) questionable FAN, (b) mild FAN, (c) moderate FAN and (d) severe FAN.