Fig. 5 | Scientific Reports

Fig. 5

From: Integrating snapshot ensemble learning into masked autoencoders for efficient self-supervised pretraining in medical imaging

Fig. 5

Effect of learning rate schedulers on Snap-MAE (Scalo) performance across 12 ECG leads. Performance of the Snap-MAE (Scalo) model is evaluated using three different learning rate schedulers: cyclic cosine, cosine, and StepLR (with gamma = 0.1 every 200 epochs). The cyclic cosine scheduler consistently outperforms the others across six metrics—AUC, AUPRC, accuracy, sensitivity, precision, and F1-score.

Back to article page