Research - Early detection of Alzheimer’s using computer vision

Introduction

Illustration of the eye

• Alzheimer’s Disease (AD) is an irreversible and progressive disease of the brain

  • Type 2 diabetes commonly leads to AD

  • Often detected at a late stage

  • Lack of cure

  • Leads to death

• No reliable early detection methods

• New research shows AD patients have less choroidal thickness and can be used as a biomarker for early detection of AD

Research Question

Can choroidal thickness be accurately measured?

Method

Figure shows OCT image of the cross-sectional view of the eye

• Choroidal space is visible in Optical Coherence Tomography (OCT)

  • OCT scans are available publicly for diabetic retinopathy

  • No scans available for AD patients publicly

• Develop algorithm to automatically identify the choroidal thickness

• Test the algorithm on scans available for diabetic retinopathy

• Identify the accuracy of the choroidal thickness

Testing

• Downloaded OCT scans from public database

• Selected 206 scan results for Normal and Diabetic Retinopathy

• Sub-selected scans that are horizontal scan from nasal to temporal region

• Total of 40 scans selected for testing

• Downloaded Manual Segmentation data

  • Experts manually marked the retinal segmentation

  • Total of 25 scans were manually scored

• Developed novel software algorithm to measure the thickness automatically

  • Developed novel software algorithm to measure the thickness automatically

• Executed the algorithm using Python 3.0 on Linux OS

• Statistically compared the results from the algorithm with the manual scoring

Novel Algorithm

Results

• 40 different scan results were tested

  • 20 Normal and 20 Diabetic Retinopathy

• Best results could be obtained with black background and contrast enhancements

• Variability in size was dependent on scan configuration

  • Size variability was minimal when the scan output was flat (Error < 0.02% p value = 0.001)

  • Size variability was significant when the scan was done at an angle (Error > 1% p value <0.005)

• The sizes detected using the algorithm are found to be matching the ground truth measurements

  • Overall Error rate: ±0.02% (p value = 0.001)

Discussion

• Several publicly available databases have OCTB scans

• Available for normal eye, diabetic retinopathy and other retinal diseases

• No scans available for Alzheimer’s patients

• Used of diabetic retinopathy for the test

• Built a novel algorithm to use the geometry of the scans to find the thickness

• Results were measured against the manual segmentation

  • Error rates were less than 0.02%

  • Less than 4 micrometers difference

References

• Gholami, Peyman, and Lakshminarayanan, Vasudevan. Optical Coherence Tomography Image Retinal Database. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2019-02-20. https://doi.org/10.3886/E108503V1

• Patton N, Aslam T, Macgillivray T, Pattie A, Deary IJ, Dhillon B. Retinal vascular image analysis as a potential screening tool for cerebrovascular disease: a rationale based on homology between cerebral and retinal microvasculature J Anat. 2005; 206:319–348. [PubMed: 15817102]

• De Silva DA, Manzano JJ, Woon FP, Liu EY, Lee MP, Gan HY, Chen CP, Chang HM, Mitchell P, Wang JJ, Lindley RI, Wong TY, Wong MC. Associations of retinal microvascular signs and intracranial large artery disease. Stroke. 2011; 42:812–814. [PubMed: 21257821]

• Ermengarda Marziani; Simone Pomati; Paola Ramolfo; Mario Cigada; Andrea Giani; Claudio Mariani; Giovanni Staurenghi, Evaluation of Retinal Nerve Fiber Layer and Ganglion Cell Layer Thickness in Alzheimer's Disease Using Spectral-Domain Optical Coherence Tomography

• Fang He ,1 Rachel Ka Man Chun ,2 Zicheng Qiu ,1 Shijie Yu ,1 Yun Shi,3 Chi Ho To, Choroid Segmentation of Retinal OCT Images Based on CNN Classifier and - Fitter