Statistical Learning for Medical Images
A compact CNN reached 99.53% accuracy on a course fundus-image split, with approximately 2.2× faster training than fine-tuned ResNet18.
Comparing Representation and Model Cost
For a medical-image course project, we compared frozen ResNet18 features with classical classifiers, end-to-end fine-tuning, and a compact CNN. I led the custom CNN and preprocessing work, testing whether a smaller model could retain useful accuracy at lower training cost.
The compact network used three convolutional stages with 16, 32, and 64 channels. Grayscale and edge-based preprocessing reduced the input representation, while the ResNet18 experiments provided a reference for accuracy and training time.
Results on the Course Dataset
On the reported three-class fundus-image split, the custom CNN reached 99.53% accuracy in 155 seconds of training; fine-tuned ResNet18 reached 100% in 348 seconds. That was approximately 2.2× faster training with a 0.47-percentage-point accuracy difference.
The comparison showed a useful cost–accuracy trade-off within this course dataset. These results describe the experimental split, rather than validation of a clinical diagnostic system.