Table 2: Comparative performance of different networks for diabetic foot ulcer detection on different intersection-over-union thresholds.
Method | IoU ≥0.5 | IoU ≥0.6 | IoU ≥0.7 | IoU ≥0.8 | IoU ≥0.9 | |||||
---|---|---|---|---|---|---|---|---|---|---|
F1 | mAP | F1 | mAP | F1 | mAP | F1 | mAP | F1 | mAP | |
FRCNN R-FCN | 0.6784 | 0.6596 | 0.6044 | 0.5618 | 0.4829 | 0.4044 | 0.2705 | 0.1487 | 0.0534 | 0.009 |
FRCNN ResNet101 | 0.6623 | 0.6518 | 0.5931 | 0.5661 | 0.4701 | 0.4087 | 0.2703 | 0.1689 | 0.0551 | 0.0112 |
FRCNN Inc-Res | 0.6716 | 0.6462 | 0.5902 | 0.5385 | 0.4592 | 0.3827 | 0.2616 | 0.1644 | 0.0483 | 0.0095 |
YOLOv5 | 0.6612 | 0.6304 | 0.5898 | 0.5353 | 0.4418 | 0.3420 | 0.2355 | 0.1175 | 0.0383 | 0.0046 |
EffDet | 0.6929 | 0.6216 | 0.6076 | 0.5143 | 0.4710 | 0.3503 | 0.2505 | 0.2167 | 0.0343 | 0.0031 |
EffDet = EfficientDet; F1 = harmonic mean of precision and recall; FRCNN Inc-Res = faster region-based convolutional neural network Inception-v2-ResNet101; IoU = intersection over union; mAP = mean average precision; R-FCN = region-based fully convolutional network; ResNet = residual neural network; YOLOv5 = You Only Look Once version 5.