García-Zapirain et al. (2018)
[26] |
Pressure ulcer
|
A CNN with two parallel branches with four convolutional layers was used to extract ROIs from HSI color space and Gaussian-smoothed images Morphological filters were applied ROIs was further processed by a CNN with four branches with eight convolutional layers; a prior model and LCDG modalities were used to make up the four input modalities
|
193 (193) |
|
Li et al. (2018)
[22] |
Wound
|
Manual skin proportion evaluation YCbCr color space; the Cr channel was used to segment skin regions Images were normalized Augmentation was applied The 13 layers taken from MobileNet were used for skin region segmentation Semantic correction was applied Ambiguous background removal
|
950 (57,000) |
|
Zahia et al. (2018)
[36] |
Pressure ulcer
|
Flashlight reflection removal Image cropping to 5 × 5 patches Patch classification was performed with a CNN with three convolutional layers Image quantification was performed by using classified patches
|
Granulation: 22 (270,762)
Necrotic: 22 (37,146)
Slough: 22 (80,636) |
Granulation:
Accuracy: 0.9201 F1: 0.9731 Necrotic:
Accuracy: 0.9201 F1: 0.9659 Slough:
Accuracy: 0.9201 F1: 0.779
|
Jiao et al. (2019)
[93] |
Burn wound
|
|
1150 |
|
Khalil et al. (2019)
[16] |
Wound
|
Image preprocessing was performed with CLAHE Statistical methods were used for color and texture feature generation NMF was used for feature map reduction GBT was used for pixel semantic segmentation
|
377 |
Accuracy: 0.96 F1: 0.9225
|
Li et al. (2019)
[23] |
Wound
|
Pixel locations were encoded Images were concatenated with location maps A MobileNet backbone was used for feature extraction A depth-wise convolutional layer was used to eliminate tiny wounds and holes
|
950 |
|
Rajathi et al. (2019)
[57] |
Varicose ulcer
|
Flashlight reflection removal Active contour segmentation was used for image segmentation Wound segments were cropped into 5 × 5 patches A CNN with two convolutional layers was used for patch classification Image quantification was performed by using classified patches
|
1250 |
|
Şevik et al. (2019)
[94] |
Burn wound
|
|
105 |
|
Blanco et al. (2020)
[59] |
Dermatological ulcer
|
Augmentation was applied Images were segmented by using super-pixels Super-pixels were classified by using the ResNet architecture Classified super-pixels were used for wound quantification
|
217 (179,572) |
|
Chino et al. (2020)
[66] |
Wound
|
Augmentation was applied UNet was used for semantic segmentation (backbone not specified) Manual input for pixel density estimation Wound measurements were returned
|
446 (1784) |
|
Muñoz et al. (2020)
[95] |
Diabetic foot ulcer
|
|
520 |
|
Wagh et al. (2020)
[55] |
Wound
|
Probabilistic image augmentation was applied DeepLabV3 based on ResNet101 was used for segmentation CRF was applied to improve segmentation accuracy
|
1442 |
|
Wang et al. (2020)
[32] |
Foot ulcer
|
Cropping and zero padding Augmentation was applied The MobileNetV2 architecture with the VGG16 backbone was used to perform semantic segmentation. Connected component labeling was applied to close small holes.
|
1109 (4050) |
|
Zahia et al. (2020)
[28] |
Pressure ulcer
|
Mark R-CNN with the ResNet50 backbone was used for image semantic segmentation A 3D mesh was used to generate the wound top view and a face index matrix Matching blocks were used for wound pose correction Face indices and wounds with corrected poses were used to measure wound parameters
|
210 |
|
Chang et al. (2021)
[96] |
Burn wound
|
|
2591 |
Accuracy: 0.913 F1: 0.9496
|
Chauhan et al. (2021)
[64] |
Burn wound
|
Augmentation was applied ResNet101 was used for feature extraction A custom encoder created two new feature vectors using different stages of ResNet101 feature maps A decoder with two convolutional layers, upsampling, and translation returned the final predictions.
|
449 |
Accuracy: 0.9336 F1: 0.8142 MCC: 0.7757
|
Dai et al. (2021)
[68] |
Burn wound
|
Style-GAN was used to generate synthetic wound images CASC was used to fuse synthetic wound images with human skin textures Burn-CNN was used for semantic segmentation
|
1150 |
|
Liu et al. (2021)
[31] |
Burn wound
|
|
1200 |
|
Pabitha et al. (2021)
[56] |
Burn wound
|
|
1800 |
Segmentation:
Accuracy: 0.8663 F1: 0.869 Classification:
Accuracy: 0.8663 F1: 0.8594
|
Sarp et al. (2021)
[49] |
Wound
|
|
13,000 |
|
Cao et al. (2022)
[18] |
Diabetic foot ulcer
|
|
1426 |
Accuracy: 0.9842 F1: 0.7696 mAP: 0.857
|
Chang et al. (2022)
[60] |
Pressure ulcer
|
|
Wound And Reepithelization: 755 (2893)
Tissue: 755 (2836) |
Wound And Reepithelization:
Accuracy: 0.9925 F1: 0.9887 Tissue:
Accuracy: 0.9957 F1: 0.9915
|
Chang et al. (2022)
[50] |
Burn wound
|
|
4991 |
Accuracy: 0.9888 F1: 0.9018
|
Lien et al. (2022)
[58] |
Diabetic foot ulcer
|
Generation of 32 × 32 image patches RGB and gray-channel 32 × 32 patches were classified with ResNet18 Classified image patches were used for wound quantification
|
219 |
|
Ramachandram et al. (2022)
[48] |
Wound
|
UNet based on a non-conventional backbone was used for wound area segmentation Wound areas were cropped and resized An encoder–decoder with the EfficientNetB0 encoder and a simplified decoder was used for wound semantic segmentation
|
Wound: 465,187Tissue: 17,000 |
Wound:Tissue:
|
Scebba et al. (2022)
[27] |
Wound
|
|
1330 |
|
