Abstract:
Diagnosis before the spread of retinal diseases is vital to prevent high level blindness
and any sort of visual impairment. Many retinal diseases can be found via the fundus
imaging which has a very important role in the observation and detection of various
ophthalmic diseases. Most previous literature has focused their approaches on
identifying individual diseases or a combination of 3-4 diseases like DR, MYA,
ARMD, MH, ODC having major research. The eye is mostly affected by more than
one underlying disease or disease marker, and uptil now most datasets had very few
classes. Recently introduced RFMiD dataset, is one of the first datasets to provide 45
different classes of ophthalmic diseases. Hence making it possible to work towards
automated multi-disease classification models which would provide great help to
highlight this issue via clinical decision support systems integrated in the medical
image diagnosis. Our work aimed to achieve higher accuracy than previous literature
and to create an CDS application from the model in understanding and predicting multi
retinal diseases. Deep learning models are excellent and have proven to be extremely
effective in solving complex image processing problems. In addition, ensemble
learning yields high generalization performance by reducing variance. Therefore, a
synthesis of transfer, ensemble, and deep learning was used in this work to create an
accurate and reliable model for multi retinal disease classification. To create the Multi
Retinal Disease Classification Model (MRDCM) we used ensemble of EfficientNetB4
and EfficientNetV2S, with our final ensemble model giving promising results. In our
evaluation, we scored an AUC of 0.973 which stands better than literature. Further our
model selection is lighter than models used in literature. The model was tested on 27
main classes of RFMiD dataset for comparison with literature.
