Our first customized Deep learning model. This repository was made to hold the contents of the ML/DL contest 2023.
Given an image we attempt to predict whether a person has either of the diseases or has a normal vision
- Diabetic Retinopathy
- Glaucoma
- Cataract
We received close to 4000 retinal images pertaining to the above four categories. The data was split into training and testing data set in the proportion 60% and 40% respectively For each of the above two categories there are four folders with the respective image sets.
We have built a custom model and not implemented the transfer learning methodology successfully
- The loaded dataset is rescaled to take values between 0 and 1
- The target image size was fixed to (150,150) to preserve maximum features
- Categorical Cross entropy was used for multiclass classification
- An early stopping with patience parameter =5 was used to prevent waste of training time by monitoring loss value
- Convolution 2D layer was added to suit the image inputs
- The final layer is a dense layer with 4 neurons(referring to number of output categories)
- All layers have an activation function of ‘relu’ while the last layer contains ‘SoftMax’ activation function