Abstract:
The recent emergence of machine learning and deep learning methods for medical image
analysis has enabled the development automated computer aided systems that can assist
ophthalmologists in making better decisions about a patient's maculopathic diseases. In particular,
Maculopathy is the field where a person can have several diseases such as Age related Macular
Degeneration (AMD) and Diabetic Macular Edema (DME). Both of the diseases can lead to
permanent loss of vision. Early detection of these diseases can reduce the level of disease severity
and this early detection of disease can facilitate the patient with better and right diagnosis as well
as proper medication. Several researches have already been conducted in order to identify and
predict AMD and DME from the images of the patients by using dictionary learning based
classification, SMO, Support Vector Machines (SVM), Principal Component Analysis (PCA),
Convolutional Neural Networks (CNN) and Wavelet Convolutional Neural Networks. All of the
research conducted by far was on scan level where the individual scan was classified as AMD,
DME or Normal.
Therefore it was a hassle for the ophthalmologist in order to observe all scans and then
deciding the category of disease, a patient is suffering. In this research we have proposed a novel
approach where CNN and RNN (LSTM) are combined to detect the retinal disease on patient level.
CNNs are used for feature extraction where LSTM is used for the classification of the image into
3 classes mainly AMD, DME and Normal Patients. After training R-CNN, we tested the model on
scan as well as on patient level. On scans level, the model is giving 99.69% accuracy while it was
tested on 647 images of Duke’s dataset. The model is tested on patient level where every scan of
the patient is categorized as AMD, DME Normal and then on the basis of most occurrences, the
model decides to which class a particular patient belongs. Model has been tested for 6 patients.
Out of 6 patients, 5 patients have been correctly classified and 1 patient has been wrongly
classified. Thus we achieve 83% accuracy of the model on patient level.
