NECST Lab Friday Talk
BEye: Software and Hardware Implementation of Retinal Vessels Segmentation for Diabetic Retinopathy Screening Tests
Lara Cavinato and Irene Fidone
MSc students in BioMedical Engineering at Politecnico di Milano
DEIB - NECSTLab Meeting Room (Building 20, basement floor)
November 11th, 2016
1.00 pm
Contact:
Marco Santambrogio
Research Line:
System architectures
Lara Cavinato and Irene Fidone
MSc students in BioMedical Engineering at Politecnico di Milano
DEIB - NECSTLab Meeting Room (Building 20, basement floor)
November 11th, 2016
1.00 pm
Contact:
Marco Santambrogio
Research Line:
System architectures
Sommario
Retinal vessels segmentation is an image processing technique that has been largely exploited in both computer vision and biomedical field.
Further improvements of such an application, in terms of performance and accuracy, could be significant in many contexts, like the automation of screening tests used to detect retinal pathologies, such as diabetic retinopathy.
To this end, this work presents an implementation of a retinal vessels segmentation algorithm using matched filtering techniques.
To enhance the performance of the proposed implementation, we accelerated it on an FPGA.
The experimental results, computed on DRIVE and STARE databases, show remarkable improvements in terms of both execution time and power efficiency (6X and 5.7X respectively) compared to the software implementation.
On the other hand, the proposed approach outperformed literature works (3X speedup) without affecting the overall accuracy and sensitivity measures.
Further improvements of such an application, in terms of performance and accuracy, could be significant in many contexts, like the automation of screening tests used to detect retinal pathologies, such as diabetic retinopathy.
To this end, this work presents an implementation of a retinal vessels segmentation algorithm using matched filtering techniques.
To enhance the performance of the proposed implementation, we accelerated it on an FPGA.
The experimental results, computed on DRIVE and STARE databases, show remarkable improvements in terms of both execution time and power efficiency (6X and 5.7X respectively) compared to the software implementation.
On the other hand, the proposed approach outperformed literature works (3X speedup) without affecting the overall accuracy and sensitivity measures.