Presentation Information

Free Paper

 

[FP-SA-53] Diabetic Retinopathy
Go back
Day
Apr 05 (Sat)
Time
15:30 - 17:00
Room
Room 23 - Imperial Hotel 3F Fuji
Topic
Retina - Medical
Chair/Coordinator
Chair)Ian Pearce、Chair)Masahito Ohji
 
 
print

FP-SA-53-10

Duration 5min, Q&A 3min

Automated Detection of Microaneurysims in Diabetic Retinopathy: Outcomes of a Novel Algorithm Based on Hidden Markov Modeling

【Speaker】
Silvestro Caputo
【Author】
Silvestro Caputo Hongying Tang Jignesh Patel Tunde Peto George Saleh


Objective/Purpose
Microaneurysms are the first sign of diabetic retinopathy (DR) and thus their detection is paramount in screening for the disease. Automated systems have begun to emerge to help with this process and each employs its own algorithms. In this study we aim to apply a new method, Hidden Markov Modeling (HMM), not previously used automated screening and evaluate its effectiveness.

Materials/Patients
Digital fundus images of diabetic patients were recruited from a population screening programme.
Two 45º fundus images (one fovea centred, one disc centred) were taken, as per UK DR Screening Programme standard. The images were anonymised and then independently analysed and scored by both the software tool and human graders as normal or abnormal (the latter based on the presence of MAs). Human graders were based at the Reading Centre at Moorfields Eye Hospital and by the HMM automated system based at Department of Computing, University of Surrey, UK.

Methods
The HMM software employed in this study, involved an algorithm designed to help identify individual MAs in a varying retinal background and was applied to DR images.
Additionally it encapsulates context dependent entities by allowing fine details to be learnt thus it improves with time.
The system, depending on the complexity of the features analysed (such as MAs, hemorrhages, exudates, blood vessels etc.) can extrapolate a different number of identifiers. These can vary from a minimum of 90 identifiers up to a maximum of 270.

Results and Conclusion
9587 digital fundus images were enrolled into the study. The automated HMM detection system of MAs showed a sensitivity of 87.97% and a specificity of 98.94%. The software performed the same as the human graders p-value = 0.0333; there was a kappa agreement coefficient (k) of 0.76136 for complete agreement between the human graders and the machine. The disagreement between the human graders and the automated system was present in 128 images of witch 99 were false positive and 29 were false negative.
Based on the result obtained in this study the use of the HMM shows promess in the Automated Detection of Microaneurysims in Diabetic Retinopathy. Further investigations will better define its application.

[ Keyword ]
Diabetic Retinopathy / Microaneurysm / Screening / Hidden Markov Modeling / Digital Fundus Images

[ Conflict of Interest ]
No

Go back