Detection of smoking status from retinal images; a Convolutional Neural Network study

Ehsan Vaghefi,David Squirrell,Natalie Walker,Song Yang,Gayl Humphrey,Sophie Hill

doi:10.1038/s41598-019-43670-0

Ehsan Vaghefi, David Squirrell + Show 4 more

Open Access

https://doi.org/10.1038/s41598-019-43670-0

Copy DOI

Journal: Scientific Reports	Publication Date: May 9, 2019
Citations: 28	License type: open-access

Affiliation: University of Auckland

Abstract

Cardiovascular diseases are directly linked to smoking habits, which has both physiological and anatomical effects on the systemic and retinal circulations, and these changes can be detected with fundus photographs. Here, we aimed to 1- design a Convolutional Neural Network (CNN), using retinal photographs, to differentiate between smokers and non-smokers; and 2- use the attention maps to better understand the physiological changes that occur in the retina in smokers. 165,104 retinal images were obtained from a diabetes screening programme, labelled with self-reported “smoking” or “non-smoking” status. The images were pre-processed in one of two ways, either “contrast-enhanced” or “skeletonized”. Experiments were run on an Intel Xeon Gold 6128 CPU @ 3.40 GHz with 16 GB of RAM memory and a NVIDIA GeForce TiTan V VOLTA 12 GB, for 20 epochs. The dataset was split 80/20 for training and testing sets, respectively. The overall validation outcomes for the contrast-enhanced model were accuracy 88.88%, specificity 93.87%. In contrast, the outcomes of the skeletonized model were accuracy 63.63%, specificity 65.60%. The “attention maps” that were generated of the contrast-enhanced model highlighted the retinal vasculature, perivascular region and the fovea most prominently. We trained a customized CNN to accurately determine smoking status. The retinal vasculature, the perivascular region and the fovea appear to be important predictive features in the determination of smoking status. Despite a high degree of accuracy, the sensitivity of our CNN was low. Further research is required to establish whether the frequency, duration, and dosage (quantity) of smoking would improve the sensitivity of the CNN.

Highlights

Cardiovascular disease continues to be the leading cause of death globally[1]
This study has shown that Computational Neural Networks can be utilised to accurately predict smoking status from retinal fundus images
In highlighting the retinal vasculature, the perivascular region and the fovea, the attention maps derived from the analysis of the contrast-enhanced image dataset, demonstrate that the Convolutional Neural Networks (CNN) has identified these regions on the images as being the most important for predicting the smoking status

Summary

Introduction

Cardiovascular disease continues to be the leading cause of death globally[1]. One of the most important risk factors in the development of cardiovascular disease is cigarette smoking[2]. Analysis of retinal images has revealed that there a number of biomarkers that are associated with increased cardiovascular risk These include vessel tortuosity and bifurcation[5] calibre[6,7,8,9,10,11,12], microvascular changes[13,14] and vascular fractal dimentions[15,16,17]. Studies used RIGA and SCES datasets and a custom CNN architecture for classification of optic-disk images to diagnose glaucoma[23]. This model was developed further to extract features and classify patients into those with or without glaucoma via a random forest classifier using a transfer learning of AlexNet[24]. The efficacy of the freely available DRIVE dataset and a custom CNN and gray-scale thresholding for segmentation of retinal vasculature has been reported[34]

Objectives

Methods

Results

Conclusion