metadata

tags:
  - fastai

Model card

Model description

A crude CV to classify retinal images. I started with a pretrained model (resnet18) and tuned it with a set of retinal images.

Intended uses & limitations

Limitations stem primarily from a naïve use of the training data. Imbalances exist in the number of training images associated with each condition. In the future, a weighting function could be applied.

Training and evaluation data

About 400 retinal images were downloaded from STATE (http://cecas.clemson.edu/~ahoover/stare/).

The model was tuned with 50 epochs. accuracy-multi: 91.4%; valid-loss: 26.0%; train-loss: 26.0%