Apaar Sadhwani, Leo Tam, Jason Su Advisors: Robert Chang, Jeff Ullman, Andreas Paepcke *All contributors are/were affiliated with Stanford University at time of their contributions. Leo Tam now works at Nvidia. Address correspondence to Apaar Sadhwani at apaars@stanford.edu or Jason Su at sujason@stanford.edu.
Motivation: Affects ~100M, many in developed, ~45% of diabetics Make process faster, assist ophthalmologist, self-help Widespread disease, enable early diagnosis/care Given fundus image Rate severity of Diabetic Retinopathy 5 Classes: 0 (Normal), 1, 2, 3, 4 (Severe) Hard classification (may solve as ordinal though) Metric: quadratic weighted kappa, (pred – real) 2 penalty Data from Kaggle ~35,000 training images, ~54,000 test images High resolution: variable, more than 2560 x 1920 Other unlabeled data from Stanford
Motivation: Affects ~100M, many in developed, ~45% of diabetics Make process faster, assist ophthalmologist, self-help Widespread disease, enable early diagnosis/care Given fundus image Rate severity of Diabetic Retinopathy 5 Classes: 0 (Normal), 1, 2, 3, 4 (Severe) Hard classification (may solve as ordinal though) Metric: quadratic weighted kappa, (pred – real) 2 penalty Data from Kaggle ~35,000 training images, ~54,000 test images High resolution: variable, more than 2560 x 1920 Other unlabeled data from Stanford
Class 0 (normal) Class 4 (severe)
Motivation: Affects ~100M, many in developed, ~45% of diabetics Make process faster, assist ophthalmologist, self-help Widespread disease, enable early diagnosis/care Given fundus image Rate severity of Diabetic Retinopathy 5 Classes: 0 (Normal), 1, 2, 3, 4 (Severe) Hard classification (may solve as ordinal though) Metric: quadratic weighted kappa, (pred – real) 2 penalty Data from Kaggle ~35,000 training images, ~54,000 test images High resolution: variable, more than 2560 x 1920 Other unlabeled data from Stanford
High resolution images Image size Batch Size Atypical in vision, GPU batch size issues 224 x 224 128 Discriminative features small 2K x 2K 2 Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples
High resolution images 0 1 Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data 2 3 Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples 4
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples Class 2
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling - Mentioned in problem statement Artifacts in ~40% images - Confirmed with doctors Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates Too few training examples
Penalty/Loss High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Class 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Hard classification non-differentiable class 0 dominates - Backprop difficult Too few training examples
Penalty/Loss High resolution images Atypical in vision, GPU batch size issues Discriminative features small Predict Grading criteria: 1 not clear (EyePACS guidelines) learn from data Incorrect labeling Class 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Hard classification non-differentiable class 0 dominates - Backprop difficult Too few training examples
Penalty/Loss High resolution images Atypical in vision, GPU batch size issues Discriminative features small Predict 2 Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Class 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Hard classification non-differentiable class 0 dominates - Backprop difficult Too few training examples
Penalty/Loss Predict High resolution images 3 Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Class 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Hard classification non-differentiable class 0 dominates - Backprop difficult Too few training examples
Penalty/Loss High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Class 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Hard classification non-differentiable class 0 dominates - Backprop difficult Too few training examples
Penalty/Loss High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Class 2.5 0 1 2 3 4 Artifacts in ~40% images Truth Optimizing approach to QWK Severe class imbalance - Squared error approximation? class 0 dominates - Differentiable Too few training examples
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates - Naïve: 3 class problem, or all zeros! Too few training examples - Learn all classes separately: 1 vs All? - Balanced while training - At test time?
High resolution images Atypical in vision, GPU batch size issues Discriminative features small Grading criteria: not clear (EyePACS guidelines) learn from data Incorrect labeling Artifacts in ~40% images Optimizing approach to QWK Severe class imbalance class 0 dominates - Big learning models take more data! Too few training examples - Harness test set?
Literature survey: Hand-designed features to pick each component Clean images, small datasets Optic disk, exudate segmentation: fail due to artifacts SVM: poor performance
Literature survey: Hand-designed features to pick each component Clean images, small datasets Optic disk, exudate segmentation: fail due to artifacts SVM: poor performance
1. Registration, Pre-processing 2. Convolutional Neural Nets (CNNs) 3. Hybrid Architecture
Registration Hough circles, remove outside portion Downsize to common size (224 x 224, 1K x 1K) Color correction Normalization (mean, variance)
Class probabilities Network in Network architecture AvgPool 7.5M parameters MaxPool (stride2) No FC layers, spatial average pooling 3 Conv layers (depth 1024) instead Transfer learning (ImageNet) MaxPool (stride2) 3 Conv layers Variable learning rates (depth 384) Low for “ ImageNet ” layers MaxPool (stride2) 3 Conv layers Schedule (depth 256) Combat lack of data, over-fitting MaxPool (stride2) Dropout, Early stopping 3 Conv layers (depth 96) Data augmentation (flips, rotation) Input Image
Recommend
More recommend
