Figure 5
From: A robust deep learning detector for sleep spindles and K-complexes: towards population norms
SS detection performance on MODA with fine-tuning after pretraining SEED on another dataset. Fine-tuning is conducted using a fraction of MODA. On the X-axis, a fraction of 0% represents no training, whereas 100% represents no restrictions in size. Shown metrics are (a) F1-score, (b) recall, (c) precision and (d) mIoU. Three cases were considered: random initialization and standard training (blue curve); pretraining on the labeled dataset MASS2-SS-E1 (the worst direct transfer, see Fig. 4) and fine-tuning on MODA (red curve); pretraining on the artificial dataset CAP-A7 and fine-tuning on MODA (green curve). The dotted line corresponds to the performance of DOSED trained on the full MODA dataset only. Outliers corresponding to the case pre-trained on MASS2-SS-E1 without fine-tuning (MODA fraction 0%) are not shown in the plots. The performance metrics for this case are F1-score 53.6%, Recall 38.0%, and Precision 92.1%.
