Authors:
João Gonçalves
;
Teresa Conceição
and
Filipe Soares
Affiliation:
Fraunhofer Portugal AICOS, Rua Alfredo Allen 455/461, 4200-135 Porto and Portugal
Keyword(s):
Convolution Neural Networks, Feature-based Machine Learning, Inter-observer Reliability, Diabetic Retinopathy.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Business Analytics
;
Cardiovascular Technologies
;
Computing and Telecommunications in Cardiology
;
Data Engineering
;
Decision Support Systems
;
Decision Support Systems, Remote Data Analysis
;
Distributed and Mobile Software Systems
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Knowledge-Based Systems
;
Mobile Technologies
;
Mobile Technologies for Healthcare Applications
;
Neural Rehabilitation
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition and Machine Learning
;
Software Engineering
;
Symbolic Systems
Abstract:
The rapid growth of digital data in healthcare demands medical image analysis to be faster, precise and, at
the same time, decentralized. Deep Learning (DL) fits well in this scenario, as there is an enormous data
to sift through. Diabetic Retinopathy (DR) is one of the leading causes of blindness that can be avoided if
detected in early stages. In this paper, we aim to compare the agreement of different machine learning models
against the performance of highly trained ophthalmologists (human graders). Overall results show that transfer
learning in the renowned CNNs has a strong agreement even in different datasets. This work also presents an
objective comparison between classical feature-based approaches and DL for DR classification, specifically,
the interpretability of these approaches. The results show that Inception-V3 CNN was indeed the best-tested
model across all the performance metrics in distinct datasets, but with lack of interpretability. In particular,
this model reaches
the accuracy of 89% on the EyePACS dataset.
(More)