Lecture 25 - Deep learning in eyecare Flashcards
Definition of artificial intelligence, machine learning and deep learning
• Artificial intelligence
- The theory and development of computer systems able to perform tasks which normally would require human intelligence.
• Machine learning
- Algorithms which improve at performing a task when exposed to more data over time.
• Deep learning
- Machine Learning algorithms are organised into multilayer neural networks which can learn from vast amounts of data.
What is deep learning?
A useful way of thinking about this compared to traditional programming is that with Traditional programming you would normally have data that you would apply rules to to provide you with answers. In deep learning the answers are used to improve the learning ability of the programme to develop rules. Instead of thinking of all the rules - what if we can apply the answers and let the machine develop the rules
Why will this be used in eye-care/optometry?
• Aging population
• 10% of NHS related to eyes
• NHS eye clinics have large numbers of false positive referrals
Applications in eyecare - DR
• Diabetic retinopathy (DR) affects more than 90 million people worldwide and is a leading cause of blindness in adults.
• Screening for DR and timely referral/treatment required to prevent blindness
• Different screening approaches around the world but all have challenges related to implementation, use of human assessors and finance. Fundus photography is an effective method of screening but there are often too few skilled professionals to read the fundus photographs and to follow up on patients.
• Using deep learning many studies have shown excellent diagnostic performance Eg Sensitivity 96.8% Specificity 87% in detection of referable DR (Abramoff et al.lOVS, 2016)
• These were not tested on real world screening programmes - issues such as ethnicity and camera type not tested
Ai trained to find Diabetic Retinopathy…
• Aravind Eye Care System in India they have tested an automated image classification system to screen millions of retinal photographs of diabetic patients.
• Google Inc.and collaborating institutions found a deep learning system trained on thousands of images can achieve physician-level sensitivity and specificity in diagnosing referable DR5
• Also identified previously unrecognized associations between image patterns in the fundus photograph and cardiovascular risk factors.
• Google is now integrating this Al technology into clinical practice in a chain of eye hospitals in India
Idx-DR System…
• This technology (related to the Google technology used in India) has been approved by the US Food and Drug Administration (FDA) for detecting moderate-to-severe DR7.
• Clinician uploads digital retinal images to cloud server
• Software then provides two possible results more than mild test is negative
Ai and OCT….
• Transformative effect on management of AMD and DMO
From a deep learning perspective there are a number of attractive qualities…
- Large and growing data set
- Macular OCTs have dense 3D structural information that can be captured consistently
- Structural detail greater than in other imaging techniques
De Fauw et al nature medicine (2018)
• Applied deep learning to 3D OCT scans from patients referred to Moorfields Eye Hospital
• Assessed performance of making referral recommendations to the performance of dinicians
• Used a different deep learning approach that predicted diagnosis and provides
referral suggestions, Training = 14884 scans
• Compared decisions to various clinicians and to a gold standard (based on future clínical outcomes)
What did De Fauw et al (2018) find?
• Suggest that their framework closely matches clinical decision making
• Allow the clinician to inspect and visualize a segmentation of the scan
• Clinician training in understanding/interpreting medical images
Ai in Glaucoma…
Algorithms have been trained:
- detect glaucomatous discs
- distinguish glaucomatous nerve fibre layer damage from normal scans
- to recognise clinically significant visual field loss
- could also use IOP and VF loss measures to improve clinical forecasting
- In future could detect progressive structural optic nerve changes in glaucoma patients
What is deep learning on fundus images capable of predicting…
• Age
• Gender
• Systolic blood pressure
• Smoking status
• Previous cardiac event
• Alzheimer’s disease
