Step 5 · Classification

ResNet50

A residual network whose skip connections let a very deep model train stably.

ImageNet weights, last 75 layers unfrozen

From the paper

What we built

ResNet-50 was initialised from pretrained weights with the last 75 layers unfrozen for domain-specific feature learning, adding Conv2D, Global Average Pooling, Dropout and Layer Normalization, fully connected ReLU layers, and a softmax output.

Transfer learning
Residual connections
Layer normalization
Dropout
AdamW

Eye photo in
ResNet50 residual baseImageNet
Conv 3×3, 512 filters → batch norm
Global average pool
Dense 1024 → dropout
Dense 512 → layer norm
Dense 256 → dropout
Softmax, 5 classes
EsotropiaExotropiaHypertropiaHypotropiaNormal

Published results

How ResNet50 did on the 81 test photos

Accuracy: 74.00%
Precision: 79.00%
Recall: 73.00%
F1 score: 74.00%

Confusion matrix for ResNet50. 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	0.93	0.00	0.00	0.00	0.07
Exotropia	0.24	0.59	0.12	0.00	0.06
Hypertropia	0.00	0.06	0.88	0.06	0.00
Hypotropia	0.00	0.25	0.12	0.62	0.00
Normal	0.00	0.18	0.06	0.06	0.71

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 (93% correct) and weakest on exotropia (59% correct).