Table 1.
Summary of different models for hematoma detection and quantification.
| Ref. | ML/DL | Algorithm/Method | Dataset Size | TBI-Related Clinical Assessment |
Performance | Main contribution |
|---|---|---|---|---|---|---|
| [36] | ML | Intensity-based, region growing algorithm | 18 | Hematoma segmentation and quantification | Mean matching ratio: 0.80 Mean correspondence ratio: 0.74 |
Proposing a semi-automated method for hematoma detection and voxel-wise volume estimation |
| [37] | ML | Multiresolution binary level set method | 15 | Hematoma segmentation and quantification | Mean Sen: 0.87 Mean precision: 0.89 |
Automated method, Using adaptive threshold as initial values for binary level set method |
| [40] | ML | FCM clustering, region based active contour | 20 | ICH segmentation | Dice coefficient:0.87 Jaccard index: 0.78 Sen: 0.79 Spec: 0.99 |
The level-set method used by active contour does not need re-initialization and converges faster. |
| [35] | ML | Thresholding for segmentation, GA-based feature selection, NN classification | _ | EDH, ICH, SDH detection | Segmentation Acc: EDH: 0.96 ICH: 0.95 SDH: 0.90 ICH detection and classification Acc: 0.90 |
Proposing independent hematoma segmentation and classification approach |
| [41] | ML | GMM, expansion maximization algorithm | 11 | ICH segmentation | _ | Developing a GMM-based model to remove skull and image’s artifacts and detect hematoma |
| [39] | ML | MDRLSE, hierarchical classifier (using pixel intensity and then SVM classifier) | 627 | ICH, SDH, EDH IVH detection | Segmentation Acc First classifier: 0.9 Second classifier: 0.94 |
Using a hierarchical classifier to first classify the IVH from the normal class and then SDH, ICH, and EDH |
| [38] | ML | DRLSE, Tree bagger classifier |
42 | SDH detection | AUC: 0.87 Sen: 0.85 Spec: 0.73 |
Proposing a method for 3D segmentation of SDH considering geometric, textural, and statistical features |
| [50] | DL | Dilated CNN | 62 | Hematoma segmentation |
Dice: 0.62 Acc: 0.95 |
Hematoma detection using an FCN model combined with dilated convolutions |
| [44] | DL | NLP | 313,318 (Qure25k dataset); 491 (CQ500 dataset) as validation set | SDH, SAH, IVH, IPH and extradural hematoma detection |
AUC ICH: 0.94 Intraparenchymal: 0.95 Intraventricular: 0.93 SDH: 0.95 Extradural: 0.97 Subarachnoid: 0.96 |
Developing a DL model to detect five different subtypes of intracranial hematoma, cranial vault fractures, mass effect, and midline shift |
| [46] | DL | Original DenseNet, attention mechanism, RNN | 329 | Acute hematoma detection | Acc: 0.818 Recall: 0.886 F1-score: 0.847 |
A combination of CNN and LSTM was used to model 3D CT labeling for brain hemorrhage detection that was benchmarked against specialist clinician. |
| [47] | DL | Custom 2D/3D mask Region of interest-based CNN | 11,021 | IPH, EDH/SDH, SAH detection andquantification | Dice: EDH/SDH: 0.86 IPH: 0.93 SAH: 0.77 Pearson correlation coefficient for volume estimation: EDH/SDH: 0.98 IPH: 0.99 SAH: 0.95 |
A custom model, extracted from the feature pyramid network [57] was implemented for hematoma segmentation, classification, and volume measurement |
| [56] | ML | Active learning to train SVM classifier, active contour | 62 | Hematoma segmentation |
Dice: 0.55 Acc: 0.97 |
The proposed model could achieve a comparable result with 5 times less labeled data compared with established ML models. |
| [42] | DL | Fuzzy-based intensifier, Autoencoder, active contour Chan-Vase model | 48 | Hematoma segmentation | Dice similarity score: 0.70±0.12 Jaccard index: 0.55±0.14 |
Implementing unsupervised NN-based method for acute hematoma segmentation |
| [43] | DL | U-net based CNN | 144 subjects from CENTER -TBI and NCT02210221 datasets [58,59] | Hematoma segmentation, volume estimation | Segmentation Dice: 0.697 Volume estimation correlation coefficient: 0.966 |
Proposing a novel Multi-view CNN with a mixed loss forhematoma segmentation and quantification |
| [48] | ML/DL | Level set method, U-net, RF | 110 | SDH segmentation and severity estimation by hematoma volume classification (0–25 cc vs. >25 cc) | Sen: 0.78 Precision: 0.76 DSC: 0.75 |
Integrating classical image processing methods and DL model to improve the average performance of hematoma detection andquantification |
| [54] | DL | CNN | 937 (CENTER -TBI and NCT02210221) [58,59]; Validation: 500 |
IPH, EAH, IVH, and perilesional oedema segmentation, detection, and volume quantification | AUC for classification of lesions greater than 0 mL: IPH: 0.87 EAH: 0.89 IVH: 0.89 Perilesional oedema: 0.89 |
Proposing a CNN-based algorithm for voxel-wise segmentation, detection, and quantification of various TBI lesions and perilesional oedema |
| [60] | DL | Multi-view CNN | 120 | ICH detection and volumetric quantification | Dice coefficient: 0.697 ICC: 0.966 |
Developing a multi-view CNN with dilated convolution and mixed loss to reduce the model sensitivity to the noise and minor shape changes. |
| [55] | DL | Kapur’s thresholding, EHO algorithm, Inception v4 network, multilayer perception | 82 | ICH detection and classification | Acc: 0.941 Precision: 0.944 Spec: 0.948 Sen: 0.926 |
Developing DL-ICH model for image preprocessing, ICH segmentation, feature extraction andclassification |
| [51] | DL | Sym-TransNet | 1357 | IPH, IVH, EDH, SDH and, SAH segmentation | Dice coefficient: IPH: 0.78 IVH: 0.68 EDH: 0.359 SDH: 0.534 SAH: 0.337 Five-class: 0.716±0.031 |
Proposing a U-net based model to detect five different hematoma subtypes |
| [45] | DL | LSTM | Training and testing: 1554, validation: 386 |
ICH detection | AUC: 0.96 | Combining CNN and RNN to form a bidirectional LSTM model for intracranial hemorrhage detection |
| [52] | DL | 3D CNN, region growing algorithm |
153 | SDH, EDH, IPH segmentation |
Median DSC SDH: 0.48 EDH: 0.71 IPH: 0.37 ICH: 0.59 |
Developing a hematoma segmentation approach using DL model with 4 different parallel pathways |