Table 3.
Mean balanced accuracy (BA), precision (P), recall (R) and area under the ROC curve (AUC) results of image classification for motion artefacts (in-vivo data set) trained on real and synthetic data sets. A 10-fold cross validation was used and each image was labelled once over all folds(mean ± std). t-aug,g-aug, m-aug, b-aug represent translational, gaussian blurring, mistriggering and breathing type augmentations respectively. b-m-aug represents a random mix of mis-tiggering and breathing artefacts to balance the data set. cs stands for cost-sensitive learning with weighted losses. All results are multiplied by 1000 and the bold font highlights the best results.
| Methods | BA | P | R | AUC |
|---|---|---|---|---|
| 3DCNN no-aug | 590 ± 85 | 713 ± 69 | 467 ± 82 | 581 ± 124 |
| 3DCNN t-aug | 679 ± 63 | 751 ± 54 | 607 ± 78 | 674 ± 87 |
| 3DCNN g-aug | 690 ± 69 | 709 ± 101 | 670 ± 91 | 685 ± 90 |
| 3DCNN m-aug | 717 ± 71 | 762 ± 78 | 673 ± 74 | 732 ± 71 |
| 3DCNN b-aug | 695 ± 62 | 703 ± 40 | 687 ± 98 | 699 ± 67 |
| 3DCNN cs | 515 ± 91 | 503 ± 57 | 520 ± 68 | 613 ± 50 |
| 3DCNN b-m-aug | 721 ± 47 | 768 ± 61 | 673 ± 40 | 735 ± 67 |
| LRCN no-aug | 629 ± 97 | 724 ± 57 | 533 ± 65 | 603 ± 71 |
| LRCN t-aug | 664 ± 55 | 722 ± 69 | 607 ± 87 | 704 ± 73 |
| LRCN g-aug | 698 ± 61 | 715 ± 73 | 672 ± 80 | 708 ± 84 |
| LRCN m-aug | 731 ± 77 | 743 ± 77 | 720 ± 128 | 826 ± 80 |
| LRCN b-aug | 719 ± 53 | 731 ± 81 | 707 ± 81 | 759 ± 93 |
| LRCN cs | 511 ± 89 | 502 ± 72 | 520 ± 48 | 608 ± 71 |
| LRCN b-m-aug | 74 ± 50 | 751 ± 84 | 733 ± 66 | 828 ± 57 |