Table 3.
Performance summary of the studies cited in figure 8, sorted chronologically. All models and validation methods have been listed, and the most important experiment’s results has been singled out for each study, with its specification in the last column, and the model used emboldened.
| Author | Year | Validation | Model | ACC | SE | SP | AUC | Specification |
|---|---|---|---|---|---|---|---|---|
| Butte et al [147] | 2011 | LOOCV | LDA | 71.43% | 47.06% | 94.64% | Temporal + Spectral SE/SP: HGG | |
| Cosci et al [139] | 2016 | 10-fold CV | KNN (K=5) | 90% | 92.8% | 88.8% | SE/SP: healthy vs other | |
| Phipps et al [4] | 2017 | LOPOCV | SVM (rbf) | 97.8% | 100% | 96.90% | balanced cancer vs other | |
| Jo et al [136] | 2018 | LOPOCV | QDA | 89% | 95% | 86.79% | 0.91 | FLIm features |
| Unger et al [5] | 2020 | LOPOCV | RF | 88.78% | 93.14% | 0.96 | tumour vs no tumour | |
| Marsden et al [109] | 2020 | LOOCV (tongue/tonsil only). Test on other with RF | RF, SVM (rbf), 1D-CNN | 86% | 87% | 0.88 | in vivo, region | |
| Walsh et al [148] | 2020 | Split train/test | LogReg, RF, SVM | 0.95 | type and activation | |||
| Wang et al [119] | 2020 | LOPOCV (3 times) train=(90% / 10%) train/validation | 2D-CNN 3D-CNN | 86.5% | 89.5% | 0.858 | 3-channel DenseNet121 | |
| Romano et al [141] | 2020 | Split train/test 75% / 25% | LDA | 73% | 88% | 67% | 0.79 | intensity and lifetime |
| Dunkers et al [149] | 2021 | Out of bag no validation | RF | 95.91% | PBS bufffer, lifetime and phasor variable | |||
| Wang et al [104] | 2021 | One patient out train=(90%10%) train/validation | Custom 3D-CNN | 84.9% | 80.95% | 0.882 | MSCD-ResNet50 preserved complexity | |
| Qian et al [150] | 2021 | train/test set train=(80% / 20%)train/validation | LogReg, RF, SVM | >85% | 0.9085 | all variable logistic regression | ||
| Marsden et al [10] | 2021 | LOPOCV | NN, RF, SVM | 100% | 93% | region level | ||
| Duran et al [138] | 2021 | 7-fold CV (train) Best model on test | NN, SVM, RF | 78% | 61% | 0.81 | ensemble | |
| Weyers et al [3] | 2022 | Pre-trained then tested | RF | 96% | 89% | 0.9 | mean over patients | |
| Neto et al [152] | 2022 | 10-fold CV | RF | 0.944 | all 2p FLIm variables | |||
| Vasanthakumari et al [178] | 2022 | LOPOCV | QDA | 88.33% | 84.21% | 90.24% | Phasor + intensity + lifetime variables | |
| Ji et al [146] | 2022 | Train: 151 Cancer/CNI +217 normal Test: Images from 48 patients | K-means as classifier | 90.90% | 100% | 0.95 | τ avg + α 2 | |
| Kröger et al [153] | 2022 | Split train/test 50% / 50% Repeat 10 000 times | Decision Tree | 88% 82% | 89% 90% | M1 MΦ M2 MΦ |
HGG: High Grade Glioma; LDA=Linear discriminant analysis; RF=Random Forest; KNN=K-Nearest Neighbors; QDA=Quadratic Discriminant Analysis; SVM=Support Vector Machine; LogReg=Logistic Regression; CNN/NN=(Convolutional) Neural-Network; LOOCV/LOPOCV/CV=(Leave-One-(Patient)-Out) Cross-Validation; ACC=Accuracy; SE=Sensitivity; SP=Specificity; AUC=(Receiver operating characteristic) Area Under the Curve.