TABLE 2.
Summary of the use of artificial intelligence for wound assessment or monitoring in diabetic foot ulcer (DFU)
| Author (year) | Type of AI | n | Wound Type | Clinical Use | Measurement Accuracy | Remote Assessment Capabilities | Challenges |
|---|---|---|---|---|---|---|---|
| Goyal et al (2020) 11 |
Faster R‐CNN with InceptionResNet V2 BayesNet Random forest InceptionV3 ResNet |
1459 | DFU | Recognition of ischemia and infection in monitoring of DFU |
Average accuracy of all models: Ischemia: 83.3% Infection: 65.8% |
Eventually transit into a remote DFU diagnosis system, but challenges need to be resolved |
Substantial datasets required for good accuracy Images provided for assessment are non‐standardised (type of camera models, devices, poses and illumination) |
| Wang et al (2019) 12 |
Associative hierarchal random field (AHRF) model Conditional random field (CRF) model |
100 | Any foot ulcer | Accurately determining boundaries of ulcer images acquired under various image acquisition conditions |
Wound recognition specificity: AHRF model – 95.5% to 99.2% CRF model 1 (conditional random field) – 89.8% to 92.7% CRF model 2 –91.1% to 98.4% Wound recognition sensitivity AHRF – 76.9% to 84.4% CRF model 1 – 61.8% to 67.4% CRF model 2 – 70.3% to 76.7% |
– | Need to expand diversity of database of real wound images, with possibility of expanding into deep learning in the future |
| Ohura et al (2019) 14 |
Convolutional Neural Networks (CNNs): SegNet LinkNet U‐Net U‐Net_VGG16 |
440 | Chronic wounds | Construct a good wound segmentation model using CNN |
SegNet: AUC 0.994, sensitivity 0.909, Specificity 0.982, accuracy 0.976 LinkNet: AUC 0.987, sensitivity 0.989, specificity 0.989, accuracy 0.972 U‐Net: AUC 0.997, sensitivity 0.993, specificity 0.993, accuracy 0.988 U‐Net_VGG16 AUC 0.998, sensitivity 0.992, specificity 0.992, accuracy 0.989 – highest accuracy |
– | May not be applicable to wound assessment and segmentation of other races (study was conducted in a Japanese population) |
| Wang et al (2017) 16 | Two‐stage support vector machine (SVM) with conditional random field (CRF) image processing | 100 | DFU | Create an automated wound detection method to determine wound area on a smartphone‐based system |
Two‐stage SVM + CRF technique: Sensitivity – 73.3% Specificity – 94.6% Computation time – 20.5 s |
– |
Need to enhance wound image database Need to recruit more clinicians to delineate wound boundaries to minimise variability |