QuickQual: Lightweight, convenient retinal image quality scoring with off-the-shelf pretrained models

QuickQual

QuickQual is a state-of-the-art method for retinal image quality scoring, using only an off-the-shelf pretrained DL model (DenseNet121) and a sklearn SVM as classifier. It can be used easily with very few lines of code, making it very lightweight and convenient to use.

Results for EyeQ test set

	Accuracy	AUC	F1	LogLoss	Kappa	QuadKappa
MCFNet	0.879993	0.95883	0.860631	0.363168	0.801693	0.895506
QuickQual	0.88627	0.968703	0.867457	0.304923	0.81072	0.901928

Code and dependencies

QuickQual only requires a pretrained DenseNet121 model and a sklearn SVM classifier. QuickQual can be used with very few lines of code:

import torch
from torchvision.transforms import functional as F
from PIL import Image
import timm
import joblib

img = Image.open('/home/justinengelmann/datastorage/kaggle/test_preprocessed/10036_left.jpeg')

model = timm.create_model('densenet121.tv_in1k', pretrained=True, num_classes=0)
model.eval().cuda()
clf = joblib.load('quickqual_dn121_512.pkl')

img = F.to_tensor(F.resize(img, 512))
img = F.normalize(img, [0.5] * 3, [0.5] * 3).cuda().unsqueeze(0)
with torch.no_grad():
    features = model(img).squeeze().cpu().reshape(1, -1)
pred = clf.predict_proba(features)  # order of class probabilities: Good, Usable, Bad
print(''.join([f'{p:.4f} ' for p in pred[0]]))  # p(Good): 0.0001 p(Usable): 0.0114 p(Bad): 0.9885 

The SVM can be downloaded from the releases page of this Github. [LINK] / [DIRECT FILE LINK]

Note: The pytorch image library (timm) can be replaced with torchvision, as the densenet121 model is also available there. See the noteboook for more details. Input images are assumed to be cropped to square without large black borders as seen in the figures and their resolution should be 512x512 or higher.

QuickQual-MEME

QuickQual MEga Minified Estimator (MEME) is an even more lightweight version of QuickQual, using the same DenseNet121 plus 10 parameters for a linear model (9 weights and one bias). QuickQual-MEME produces a one-dimensional image quality score instead of predictions for three separate classes.

QuickQual-MEME does not even require sklearn classifier, the linear model with its 10 parameters is contained in the code below:

import torch
from torchvision.transforms import functional as F
from PIL import Image
import timm

img = Image.open('/home/justinengelmann/datastorage/kaggle/test_preprocessed/10036_left.jpeg')

model = timm.create_model("densenet121.tv_in1k", pretrained=True, num_classes=0)
model.eval().cuda()

w = torch.tensor([-1411.32, 517.09, 342.41, -707.9,
                  1442.09, -23.25, -541.64, -8.44, 5.44])
b = torch.tensor([5.18])

img = F.to_tensor(F.resize(img, 512))
img = F.normalize(img, [0.5] * 3, [0.5] * 3).cuda().unsqueeze(0)

with torch.no_grad():
    feats = model(img).squeeze().cpu().reshape(1, -1)
feats = feats[:, [71, 109, 121, 53, 55, 123, 29, 133, 84]]
pred = torch.sigmoid(feats @ w + b)
print(f'Predicted p(bad): {pred.item():.4f}')  # Predicted p(bad): 0.9386

Acknowledgements

We thank Huazhu Fu and colleagues for releasing the EyeQ dataset and MCFNet code and results. This is an amazing contribution to the field and without their work, QuickQual would not have been possible. Check out their repo: https://github.com/HzFu/EyeQ