Table 4.

Summary of studies on chronic wound classification.

Reference	Subject and Classes (Each Bullet Represents a Different Model)	Methodology	Original Images (Training Samples)	Results
Goyal et al. (2018) [51]	Diabetic foot ulcer Wound/Non-Wound	Cropping and resizing Augmentation was applied A 14-layer DFUNet CNN with parallel convolutions was used for binary classification	344 (22,605)	Accuracy: 0.896 F1: 0.915
Cirillo et al. (2019) [63]	Burn wound Superficial/Pre-Intermediate/Post-Intermediate/Full/Skin/Background	224 × 224 image patches were cropped Augmentation was applied ResNet101 was used for classification	23 (676)	Accuracy: 0.9054
Zhao et al. (2019) [53]	Diabetic foot ulcer Wound Depth (0–4) Granulation Tissue Amount (0–4)	256 × 256 image patches were cropped Sharpening was applied Augmentation was applied Bi-CNN was used for classification. VGG16 backbones were used for depth and granulation paths.	1639	Wound Depth: Accuracy: 0.8336 F1: 0.83024 Granulation Tissue Amount: Accuracy: 0.8334 F1: 0.82278
Abubakar et al. (2020) [52]	Burn wound Wound/Non-Wound	Cropping and normalization Augmentation was applied ResNet50 was used for image classification	1900	Accuracy: 0.964
Alzubaidi et al. (2020) [79]	Diabetic foot ulcer Wound/Non-Wound	Cropping and resizing Augmentation was applied A 58-layer DFU_QUTNet was used for feature extraction Extracted features were used in an SVM classifier	754 (20,917)	F1: 0.945
Alzubaidi et al. (2020) [47]	Diabetic foot ulcer Wound/Non-Wound	224 × 224 image patches were cropped A CNN was trained by using a dataset of the same domain A CNN with 29 convolutional layers (parallel and simple) was used for classification	1200 (2677)	F1: 0.976
Chauhan et al. (2020) [62]	Burn wound Severe/Low/Moderate	Augmentation was applied by using label-preserving transformations ResNet50 was used for body part classification Body-part-specific ResNet50 was used for feature extraction, and an SVM was used to classify burn severity	141 (316)	Accuracy: 0.8485 F1: 0.778
Goyal et al. (2020) [40]	Diabetic foot ulcer Ischemia/Non-Ischemia Infection/Non-Infection	Cropping and resizing Natural augmentation was applied InceptionV3, ResNet50, and InceptionResNetV2 were used for feature extraction An SVM classifier used extracted bottleneck features for classification	Ischemia: 1459 (9870) Infection: 1459 (5892)	Ischemia: Accuracy: 0.903 F1: 0.902 MCC: 0.807 Infection: Accuracy: 0.727 F1: 0.722 MCC: 0.454
Wang et al. (2020) [54]	Burn wound Shallow/Moderate/Deep	Images were cropped and resized Augmentation was applied ResNet50 was used for classification	484 (5637)	F1: 0.82
Rostami et al. (2021) [33]	Wound Background/Normal Skin/Various ulcer/Surgical Wound	Augmentation was applied The AlexNet classifier was used for whole-image classification 17 image patches were cropped The AlexNet classifier was used for patch classification An MLP classifier used the results of the two AlexNet classifiers to make a final decision	400 (19,040)	Accuracy: 0.964 F1: 0.9472
Xu et al. (2021) [45]	Diabetic foot ulcer Ischemia/Non-Ischemia Infection/Non-Infection	The DeiT vision transformer was used for feature extraction An MLP reduce feature map dimensionality for storage in a knowledge bank Classification was based on cosine similarity between test images and stored knowledge	Ischemia: 9870 (9870) Infection: 5892 (5892)	Ischemia: Accuracy: 0.909 F1: 0.903 Infection: Accuracy: 0.78 F1: 0.782
Al-Garaawi et al. (2022)	Diabetic foot ulcer Wound/Non-Wound Ischemia/Non-Ischemia Infection/Non-Infection	Augmentation was applied Local binary patterns were extracted and converted into 3D representations RGB and LBP were summed The sum was fed into a 12-layer CNN.	Wound: 1679 (16,790) Ischemia: 9870 (9870) Infection: 5892 (5892)	Wound: Accuracy: 0.93 F1: 0.942 Ischemia: Accuracy: 0.99 F1: 0.99 Infection: Accuracy: 0.742 F1: 0.744
Al-Garaawi et al. (2022) [43]	Diabetic foot ulcer Wound/Non-Wound	GLCM, Hu moment, color histogram, SURF, and HOG feature extraction A CNN with four residual blocks was used for feature extraction A logistic regression classifier was used to classify images	1679 (1679)	Accuracy: 0.9424 F1: 0.9537
Alzubaidi et al. (2022) [30]	Diabetic foot ulcer Wound/Non-Wound	Cropping and resizing Augmentation was applied A CNN with 29 convolutional layers (parallel and simple) was used for classification	3288 (59,184)	F1: 0.973
Anisuzzaman et al. (2022) [29]	Wound Background/Normal Skin/Various ulcers/Surgical Wound	Images were cropped and locations were labeled Data augmentation was applied The VGG19 classifier was used for wound image classification An MLP was used for wound location classification The results of previous classifiers were concatenated An MLP was used for the final classification	1088 (6108)	F1: 1
Das et al. (2022) [80]	Diabetic foot ulcer Wound/Non-Wound	Cropping and resizing A CNN with three parallel convolutional layers was used for classification	397 (3222)	Accuracy: 0.964 F1: 0.954
Das et al. (2022) [81]	Diabetic foot ulcer Wound/Non-Wound Ischemia/Non-Ischemia Infection/Non-Infection	Grayscale image conversion Gabor and HOG feature extraction GoogLeNet was used to extract features A random forest classifier was used to classify images	Wound: 1679 (1679) Ischemia: 9870 (9870) Infection: 5892 (5892)	Wound: Accuracy: 0.88 F1: 0.89 Ischemia: Accuracy: 0.92 F1: 0.93 Infection: Accuracy: 0.73 F1: 0.76
Das et al. (2022) [44]	Diabetic foot ulcer Ischemia/Non-Ischemia Infection/Non-Infection	Res4Net (network with four residual blocks) was used for ischemia classification Res7Net (network with seven residual blocks) was used for infection classification	Ischemia: 9870 (9870) Infection: 5892 (5892)	Ischemia: Accuracy: 0.978 F1: 0.978 MCC: 0.956 Infection: Accuracy: 0.8 F1: 0.798 MCC: 0.604
Liu et al. (2022) [41]	Diabetic foot ulcer Ischemia/Non-Ischemia Infection/Non-Infection	Augmentation was applied EfficientNet networks were used for classification (B1 for ischemia, B5 for infection)	Ischemia: 2946 (58,200) Infection: 2946 (58,200)	Ischemia: Accuracy: 0.9939 F1: 0.9939 Infection: Accuracy: 0.9792 F1: 0.9792
Venkatesan et al. (2022) [78]	Diabetic foot ulcer Wound/Non-Wound	Natural dataset transformation using CLoDSA SMOTE was applied for class balancing A 22 layer CNN with parallel convolutions was used for classification	1679 (18,462)	Accuracy: 1 F1: 1 MCC: 1
Yogapriya et al. (2022) [17]	Diabetic foot ulcer Infection/Non-Infection	Augmentation was applied A custom CNN DFINET is used for image classification	5892 (29,450)	Accuracy: 0.9198 F1: 0.9212 MCC: 0.84