Table 2.
Comparative analysis of various classification approaches.
| Acc (%) | |
|---|---|
| MobilenetV2 (raw) | 55.8 |
| MobilenetV2 (raw) (CA) | 50.5 |
| TA base MobilenetV2 | 83.5 |
| TA base MobilenetV2 (CA) | 88.8 |
| 4x TA MobilenetV2 | 90.3 |
| 4x TA MobilenetV2 (CA) | 92.2 |
| PGGAN+kmeans base MobilenetV2 | 84.5 |
| PGGAN+kmeans base MobilenetV2 (CA) | 88.3 |
| 4x PGGAN+kmeans MobilenetV2 | 89.8 |
| 4x PGGAN+kmeans MobilenetV2 (CA) | 92.7 |
| CycleGAN+kmeans base MobilenetV2 | 83.5 |
| CycleGAN+kmeans base MobilenetV2 (CA) | 88.3 |
| 4x CycleGAN+kmeans MobilenetV2 | 87.4 |
| 4x CycleGAN+kmeans MobilenetV2 (CA) | 91.7 |
Acc (accuracy), kmeans (k = 1, removes PGGAN generative images. K = 2, removes CycleGAN generative images), base (augmentation applied to increase normal set size same as acute set size so as to mitigate the imbalance). TA, traditional augmentation; CA, channel attention. Bold means highest accuracy.