Step 5 · Classification

VGG19

A deep, uniform stack of 3×3 convolutions, pretrained on ImageNet and fine-tuned in stages.

ImageNet weights, progressive unfreezing

From the paper

What we built

VGGNet-19 used Flatten layers, Batch Normalization and Dropout, with adjusted fully connected layers. A progressive unfreezing approach gradually unfroze deeper layers for task-specific features, while the outer layers stayed frozen to keep their learned representations.

Transfer learning
Progressive unfreezing
Batch normalization
Dropout
L2 regularization

Eye photo in
VGG19 convolutional base16 conv layers, ImageNet
Flatten
Dense 512 → batch norm → dropout
Dense 256 → batch norm → dropout
Dense 128 → batch norm → dropout
Softmax, 5 classes
EsotropiaExotropiaHypertropiaHypotropiaNormal

Published results

How VGG19 did on the 81 test photos

Accuracy: 66.67%
Precision: 67.45%
Recall: 66.67%
F1 score: 67.00%

Confusion matrix for VGG19. Rows are the true class, columns the predicted class; each cell is the share of that true class's test photos.
True ↓	Esotropia	Exotropia	Hypertropia	Hypotropia	Normal
	Predicted class →
Esotropia	0.80	0.13	0.00	0.00	0.07
Exotropia	0.06	0.47	0.18	0.24	0.06
Hypertropia	0.06	0.25	0.62	0.06	0.00
Hypotropia	0.06	0.12	0.06	0.69	0.06
Normal	0.00	0.18	0.06	0.00	0.76

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 esotropia (80% correct) and weakest on exotropia (47% correct).