Strabismus

Step 5 · Classification

EfficientNet-B7

The largest EfficientNet, which scales depth, width and resolution together. The best performer.

ImageNet weights, last 30 layers unfrozen

From the paper

What we built

EfficientNet-B7 was selected as the final model because it outperformed the others. It integrated Conv2D layers, Global Average Pooling, weight-decay regularization and an adaptive learning-rate optimizer, with the last 30 layers unfrozen to refine its representations on this dataset.

  • Transfer learning
  • Compound scaling
  • Weight decay
  • Mixed precision
  • Adaptive learning rate
  1. Eye photo in
  2. EfficientNet-B7 basecompound-scaled, ImageNet
  3. Conv 3×3, 512 filters → batch norm
  4. Global average pool
  5. Dense 1024 → dropout
  6. Dense 512 → layer norm
  7. Dense 256 → dropout
  8. Softmax, 5 classes
  9. EsotropiaExotropiaHypertropiaHypotropiaNormal

Published results

How EfficientNet-B7 did on the 81 test photos

Accuracy
84.00%
Precision
85.00%
Recall
83.00%
F1 score
84.40%
Confusion matrix for EfficientNet-B7. Rows are the true class, columns the predicted class; each cell is the share of that true class's test photos.
Predicted class →
True ↓EsotropiaExotropiaHypertropiaHypotropiaNormal
Esotropia0.800.070.000.070.07
Exotropia0.060.760.000.060.12
Hypertropia0.060.000.940.000.00
Hypotropia0.000.060.060.750.12
Normal0.000.000.000.060.94
01 — share of each true class; the diagonal is correct predictions

Each row is a true class; the numbers show where its test photos ended up. The diagonal is the share classified correctly.

It was strongest on hypertropia (94% correct) and weakest on hypotropia (75% correct).