Ensemble deep learning model for the detection of retinal diseases
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The human sense of vision plays a vital role in maintaining
newlineproductivity and independence. Unfortunately, a significant number of people
newlineare affected by retinal diseases that, if not treated promptly, can lead to
newlineblindness. Clinical experts traditionally analyze ophthalmic images, but the
newlinechallenges arise from the complexity, volume, and variations in pathology.
newlineA substantial challenge revolves around the detection of physiological
newlinechanges within the human body. Currently, these irregularities are done
newlinemanually which proves tedious and time-consuming due to the complex
newlinenature of the methods used. To overcome these problems, there is a rising
newlineinterest in computer-assisted assessment and diagnosis of these images. This
newlineResearch aim to develop a cost-effective and robust approach that can
newlineefficiently identify retinal diseases in large populations without the need for
newlinehuman intervention. Accurate early detection of retinal diseases can prevent
newlineirreversible vision loss and improve patient outcomes.
newlineThe first module of the proposed research work is to build an
newlineensemble deep learning model comprising three pre-trained convolutional
newlineneural networks such as ResNet50, VGGNet19, and Inception-V3 for early
newlineidentification of glaucoma. The proposed algorithm is applied on four distinct
newlinepublicly available datasets such as RIM-ONE, DRISHTI-GS, DRIONS-DB,
newlineand HRF which are combined together. The heads of all three deep nets are
newlinemodified which acts as a binary classifier to give a final output as normal or
newlineglaucomatous image. The results demonstrate an accuracy of 98.58%,
newlinespecificity of 98.80%, sensitivity of 97.43%, precision of 96.02%, F1 score of
newline96.72%, and an AUC of 0.98 for the combined dataset. The evaluation of
newlineiv
newlineperformance shows a remarkable accuracy rate which are also achieved across
newlineindividual datasets: 98.67%, 97.71%, 97.22%, and 97.50% for
newlineRIM-ONE, DRIONS-DB, HRF, and DRISHTI-GS, respectively. The
newlineensemble framework exhibits a significant advantage over the state-of-the-art
newlinetechniques, showcasing a performan