Table 1.
Summary of different ML-based burn classification studies using imaging data as inputs to the ML algorithms. Modality, data processing techniques, ML classifiers, validation procedures, and reported accuracy values are shown in the columns of the table.
| Data modality | Studies | Pre-processing | ML classifier | Validation methods | Accuracy |
|---|---|---|---|---|---|
| Digital color |
17–52 |
E.g., , , ; texture analysis |
E.g., SVM, LDA, KNN, deep CNN |
E.g., k-fold CV; separate test set |
80.9% +/− 6.4% without deep learning; 86.2% +/− 9.8% with deep learning (see Fig. 7) |
| Multispectral | 53 | Outlier detection | SVM, KNN | tenfold CV | 76% |
| 54 | LDA, QDA, KNN | 68%–71% | |||
| 55 | CNN | Sensitivity = 81%; PPV = 97% | |||
| Hyperspectral | 56 | Denoising | Unsupervised segmentation | Comparison between segmentation and histology | Not reported |
| Multispectral SFDI |
57
|
Calibrated reflectance |
SVM |
tenfold CV |
92.5% |
| Digital color + multispectral |
58,59 | Texture analysis, mode filtering | QDA | twelvefold CV (in Ref. 59) | 78% |
| 60 | QDA + k-means clustering | 34-fold CV | 24% better than QDA alone for identifying non-viable tissue | ||
|
61
|
Outlier removal using Mahalanobis distance |
|
|
|
|
| OCT |
62 | OCT and pulse speckle imaging | Naïve Bayes classifier | ROC AUC = 0.86 | |
| 63 | A-line, B-scan, and phase data | Multilevel ensemble classifier | tenfold CV | 92.5% | |
|
64
|
Eight OCT parameters |
Linear model classifier |
Test set |
91% |
|
| Ultrasound |
65 (ex vivo) | Texture analysis | SVM and kernel Fisher | |
93% |
|
66 (in situ postmortem) |
B-mode ultra sound data |
Deep CNN |
99% |
||
| Thermography |
67 | Thermography and multispectral | CNN pattern recognition | Training, validation, and test sets | Precision = 83% |
|
68
|
Temperature difference relative to healthy skin |
Random forest |
Training and validation sets |
85% |
|
| Blood Flow |
69
|
LSI |
CNN |
|
>93% |
| Terahertz Imaging | 70,71 | Wavelet denoising, Wiener deconvolution | SVM, LDA, Naïve Bayes, neural network | Fivefold CV | ROC AUC = 0.86-0.93 |
| 72 | Permittivity | Three-layer fully connected neural network | Fivefold CV | ROC AUC = 0.93 |
Note: PPG, photoplethysmography; OCT, optical coherence tomography; LSI, laser speckle imaging; SVM, support vector machine; LDA, linear discriminant analysis; KNN, K-nearest neighbors; QDA, quadratic discriminant analysis; CNN, convolutional neural network; CV, cross-validation; ROC, receiver operating characteristic; AUC, area under curve.