Skip to main content
Fig. 10 | BMC Medical Informatics and Decision Making

Fig. 10

From: On the analysis of data augmentation methods for spectral imaged based heart sound classification using convolutional neural networks

Fig. 10Fig. 10

PR curves for Model 0 (a), Model 1 (b), Model 2 (c), Model 3 (d, e), Model 4 (fh), Model 5 (i), Model 6 (j), Model 7 (k). Comparison of the PR curve for Model 0, trained on real data only (a); PR curve for Model 1, trained on Mel-Spectrograms of real plus pitch shifted and time stretched/compressed heart sounds (b); PR curve for Model 2, trained on Mel-Spectrograms of real plus noise injected heart sounds (c); PR curves for Models 3.1 and 3.2, trained on real and horizontally flipped Mel-Spectrograms (d), and real and vertically flipped Mel-Spectrograms (e); the PR curves for Model 4.1, 4.2, and 4.3, trained on real and saturation/value transformed images (f), real and multi-color transformed Mel-Spectrograms (g), real and PCA color augmented Mel-Spectrograms (h); the PR curve for Model 5, trained on real and frequency/time masked Mel-spectrograms (i); the PR curve for Model 6, trained on real and horizontally flipped/PCA augmented Mel-spectrograms (j); and the PR curve for Model 7, trained on real and horizontally flipped/SV perturbed Mel-spectrograms (k). The dotted red line represents the no-discrimination line

Back to article page