Introduction
Brain tumor and stroke lesions
Brain imaging modalities
Positron emission tomography
Computed tomography
Magnetic resonance imaging
Diffusion weighting imaging
Evaluation and validation
Publicly available datasets
Datasets | Description | Sequences | Number of slices (images) |
---|---|---|---|
BRATS series | BRATS 2012 challengeimage (Patients) dataset (05 LGG, 10 HGG) casesSyntheticdataset (04 LGG,11HGG) cases | T1weighted, T1C weighted, T2 weighted and Flair | 240 × 240 × 155 |
BRATS 2012 Trainingimage dataset (10 LGG, 20 HGG) casesSyntheticdataset (25 LGG, 25 HGG) cases | |||
BRATS 2013 challenge30 Subjects | |||
BRATS 2013 Leaderboard, 25 Subjects | |||
2014 challenge, 400 Subjects | |||
2015 challenge, 274 Subjects | |||
2016 challenge, Training cases of BRATS 2015 | |||
BRATS 2017 challenge, 285 Subjects | |||
BRATS 2018 challenge, 191 Subjects | |||
BRATS 2019 challenge, 22,087 training and 22,087 testing slices | |||
BRATS 2020 challenge, 25,962 training and 25,962 testing slices | |||
Harvard [58] | 65 tumor and 35 non-tumor images | T2 weighted | \(256 \times 256\) (100 images) |
RIDER[59] | 126 Subjects | T1 weighted, T2 weighted, and Flair | \(256 \times 256\) 126 cases |
ISLES 2015 | 64 Subjects | SISS- ISLES DWI, T1 weighted, T2 weighted, Flair SPES-ISLES CBF, CBV, DWI, T1C weighted, T2 weighted, Tmax, TTP | SISS- ISLES \(230 \times 230 \times 154\) (154 slices in each case) SPES-ISLES \(230 \times 230 \times 154\) (154 slices in each case) |
ISLES 2016 | 75 Subjects | MTT, rCBV, relative rCBF, Tmax, TTP | \(192 \times 192 \times 19\) (19 slices in each case) |
ISLES 2017[60] | 57 Subjects | PWI, ADC, MTT, rCBV, rCBF, Tmax, TTP | \(192 \times 192 \times 19\) (19 slices in each case) |
Performance metrics
Preprocessing
-
Conventional methods.
-
Machine learning methods.
-
Different inhomogeneities related to MRI noise have shading artifacts and partial volume effects.
References | Year | Segmentation methods | Datasets | Outcomes |
---|---|---|---|---|
[79] | 2005 | Hybrid level set (HLS) segmentation | 10 patients | 79.12 to 93.25% matching (PM) |
[78] | 2013 | A semi-automatic method based on individual and population information | 137 clinical images | 94.1 ± 3.0 DSC |
[80] | 2015 | Fully automated generative method | BRATS (2013 Challenge, 2013 Leaderboard, 2015 Challenge) | 0.87 DSC complete, 0.82 core, 0.70 enhance on BRATS 2013 Challenge, 0.83 complete, 0.71 core, 0.54 enhance on BRATS 2013 Leaderboard, 0.81 DSC complete, 0.68 core DSC, 0.65 DSC enhance on BRATS 2015 Challenge |
[81] | 2016 | Expectation maximization | SPES and SISS 2015 | 0.78 ± 0.08 DSC on SPES and 0.53 ± 0.26 DSC on SISS |
[82] | 2017 | Otsu algorithm | BRATS 2013 Synthetic | 0.93 ± 0.04 DSC on HG, 0.90 ± 0.02 DSC on LG 0.87 ± 0.06 Jaccard Index on HG, 0.82 ± 0.04 Jaccard Index on LG |
[83] | 2017 | Non-negative matrix factorization (NMF) | 21 HGG patients | 0.80 complete DSC, 0.74 core DSC and 0.65 active DSC tumor |
[84] | 2017 | HCSD | BRATS2012 Challenge | 0.9102 ± 0.0627 DSC, 0.9501 ± 0.0518 SE, 0.9980 ± 0.0023 SP |
[85] | 2018 | Improved thresholding method | Harvard and Private collected images | 0.948 Jaccard index on clinical and 0.961 Jaccard index on Harvard |
[86] | 2018 | Novel saliency method | BRATS 2013 Challenge | 0.86 ± 0.06 HG DSC, 0.85 ± 0.07 LG DSC |
[87] | 2018 | BA and RG | BRATS 2015 Challenge | 0.8741 Jaccard index, 0.9036 DSC, 0.9827 sensitivity, 0.9772 specificity, 0.9753 accuracy and 0.9585 precision |
[88] | 2019 | EM and FODPSO | 192 MRI scan | 0.93.4 ACC |
[89] | 2019 | Adaptive threshold and morphological operations | 1340 Clinical MR images | 0.85 DSC, 0.89 Jaccard index |
[90] | 2020 | 3D semantic segmentation | BRATS 2019 challenge | 0.826 enhance, 0.882 complete, 0.837 core tumor |
[91] | 2021 | CNN model | FLAIR, (T1T1C, and T2) weighted | 0.957 ACC |
Conventional methods
-
Thresholding methods.
-
Region growing methods.
-
Watershed methods.
Segmentation
Thresholding methods
Region growing (RG) methods
Watershed methods
Feature extraction methods
References | Year | Extracted features | Dataset | Results |
---|---|---|---|---|
[154] | 2013 | Gabor-like multiscale texton features | BRATS2012 | 0.73 DSC |
[148] | 2015 | GLCM features | 120 MR images | 0.817 similarity index, 0.817 overlap function, 0.182 extra function, and 0.817 PPV |
[15] | 2017 | Shape, texture, and intensity features | Harvard, RIDER, Private collected images | 0.79 ACC on the cubic kernel of SVM (Private collected images), 0.96 ACC on the cubic kernel of SVM (RIDER), 0.87 ACC on the cubic kernel of SVM (Harvard) |
[144] | 2017 | 371 texture and intensity features | Harvard | 0.9334 ACC |
[145] | 2017 | Shape descriptor | 90 MR images | 0.9889 ACC |
[146] | 2017 | Multi-fractal features | Harvard | 98.01% ± 0.07 ACC, 1.00 SE, and 94.78% ± 0.02 SP |
[147] | 2017 | GLCM, GLGCM, GLCCM and Tamura features | 62 patients | 0.7581 ACC, 0.8122 AUC |
[14] | 2018 | GWF, HOG, LBP SFTA features | 2012 Image, 2013 challenge, 2015 challenge [BRATS], ISLES 2015 | 0.98 SE on 2012 Image, 0.98 SE, on 2013 challenge, 0.98 SE, on 2015 challenge, 1.00 SE, on ISLES 2015 |
[11] | 2019 | LBP and GWF and fusion of both LBP and GWF features | BRATS2013 Challenge, BRATS 2015 Challenge Private collected images | 1.00 SP on BRATS 2013, 0.90 SP on Fused feature vector (ensemble classifier) on BRATS 2015, 0.83 SP, 0.91 on Fused feature vector (ensemble classifier) on private collected images |
[150] | 2019 | GLCM features | 105 MR images | 0.9882 ACC, 1.00 SE, 0.9783 SP, and 1.17 Error rate |
[155] | 2020 | Stochastic texture features | 9 Patients of BRATS 2015 | 0.852 ± 0.063 complete, 0.812 ± 0.074, 0.851 ± 0.093 enhance |
[156] | 2020 | CNN, LBP, and HOG features | BRATS 2015 | 0.81 complete, 0.76 core and 0.73 enhance |
[157] | 2021 | (PCA), entropy, mean, and wavelet transform | BRATS 2015 | 0.96 ACC |
Feature selection methods or feature selection/reduction methods
References | Year | Methods | Classifiers | Datasets | MRI sequences | Classes | Results |
---|---|---|---|---|---|---|---|
[158] | 2011 | Wavelet transform, PCA, scaled conjugate gradient (SCG) | Neural network (NN) | Harvard | T2 weighted | Normal or abnormal | 1.00 ACC |
[159] | 2012 | PCC, PCA, and ICA | SVM | Private data | T2 weighted and T1 weighted | Low- and high-grade glioma | 0.82 ACC at PCC, 0.85 ACC at PCA, 0.79 ACC at ICA |
[160] | 2013 | Wavelet, shape, texture, and boundary features, ICA | Ensemble classifier (SVM,KNN, and ANN) | Private images collected from ShirdiSai Cancer Hospital | T1 weighted and T2 weighted | Benign and malignant | 0.99 ACC, 1.00 SE, 0.98 SP |
[45] | 2014 | Feedback PCNN (FPCNN), wavelet transform, PCA | ANN | 14 normal and 87 abnormal images | T2 weighted | Normal and abnormal | 0.99 ACC |
[161] | 2014 | Multi-dimensional co-occurrence matrix | ELM-IPSO | 35 clinical routine cases | T1 weighted and T1C weighted | Low and high-grade glioma | 0.99 ACC, 0.95 SP and 0.98 SE |
[162] | 2015 | N4ITK, histogram matching | Random decision forest (RDF) | BRATS 2013 Challenge | FLAIR, (T1T1C, and T2) weighted | Complete, core and enhance | 0.86 complete.0.71 core,0.73 enhance DSC on LGG,0.76 complete,0.58 core, 0.16 enhance DSC on HGG |
[163] | 2015 | Intensity, geometry, and asymmetry | RF | BRATS 2013 Challenge | FLAIR, T1, T1C, and T2 (weighted) | Complete, core and enhance | 0.87 complete, 78 core, 74 enhance DSC, 85 complete,74 core, 69 enhance PPV, 89 complete,88 core, 83 enhance SE |
[164] | 2016 | 2D-DWT, Wavelet-energy | SVM | Harvard | T2 (weighted) | Normal and abnormal | 0.97 ACC, 0.99 precision,0.92 SP, 0.98 SE |
[82] | 2017 | Multi-level Otsu threshold, HOG-TOP | RF | BRATS 2013 synthetic | Flair | Complete, core and enhance | 0.93 ± 0.04 HG DSC, 0.90 ± 0.02 LG DSC |
[145] | 2017 | HMI | Twin SVM | Harvard | T2 weighted | Normal and abnormal | 0.98 ACC, 0.99 SE, 0.92 SP |
[146] | 2017 | Directional spectral distribution | SVM | Harvard | T2 weighted | Healthy and glioma | 1.00 SE |
[165] | 2018 | Stationary wavelet entropy (SWE) | Kernel support vector machine (KSVM) | Harvard | T2 weighted | Normal and abnormal | 0.98 ACC, 0.99 precision, 0.96 SP, 0.98 SE |
[166] | 2018 | Fractional Sobel filter, statistical features | SVM | BRATS 2013 challenge | T1 weighted and Flair | Tumor and healthy tissues | 0.98 ACC, 0.86 SE, 0.98 SP |
[27] | 2020 | Stacked sparse autoencoder (SSAE) | Softmax | 2012, 2013, 2014, and 2015 BRATS | FLAIR, T1, T1C,T2 weighted | Normal and abnormal | 100% on 2012, 90% on 2012 synthetic, 95% on 2013, 100% on Leaderboard 2013, 97% 2014 and 95% |
[167] | 2021 | CNN | SVM | Local data | FLAIR, T1, T1C, T2 weighted | Grade I, II, III and IV | 100% |
Classification methods
Recent trends in medical imaging to detect malignancy
Deep learning methods
References | Year | DL model | Datasets (BRATS 2012–2019) | Types of tumor | Performance measures (DSC) | ||
---|---|---|---|---|---|---|---|
Complete | Core/non-enhanced | Enhanced | |||||
[225] | 2014 | CNN | 2013 | Glioma | 83.7 ± 9.4 | 73.6 ± 25.6 | 69.0 ± 24.9 |
[226] | 2015 | CNN | 2014 | 0.81 ± 15 | 0.79 ± 13 | 0.81 ± 11 | |
[227] | 2016 | CNN | 2013 | 0.88 | 0.83 | 0.77 | |
[193] | 2017 | Input cascade CNN | 2012 | 0.81 | 0.72 | 0.58 | |
2013 Leaderboard | 0.84 | 0.71 | 0.57 | ||||
2013 Challenge | 0.88 | 0.79 | 0.73 | ||||
[219] | 2017 | U-Net CNN | 2015 | 0.86 | 0.86 | 0.65 | |
[218] | 2017 | DeepMedic + CRF | 2015 | 0.84 | 0.67 | 0.62 | |
ISLES 2015 | 0.66 DSC | – | – | ||||
[228] | 2018 | FCNN + CRF | (2013 Challenge | 0.85 | 0.83 | 0.74 | |
2013 Leaderboard | 0.88 | 0.84 | 0.77 | ||||
2015 Challenge) | 0.82 | 0.72 | 0.62 | ||||
[229] | 2018 | DNN (ILinear nexus architecture) | 2013 | 0.87 | 0.89 | 0.92 | |
2015 | 0.86 | 0.87 | 0.90 | ||||
[220] | 2019 | Dual-force CNN | 2015 | 0.83 | 0.67 | 0.63 | |
2017 | 0.87 | 0.73 | 0.69 | ||||
[221] | 2019 | WRN-PPNet | 2015 | 0.94 | – | – | |
2018 | 0.91 | ||||||
[230] | 2021 | YOLOv2 | 2018, 2019, 2020 | 0.90 | – | – | |
[231] | 2020 | 3D U-Net and DeepMedic | 2017 | 0.90 | 0.81 | 0.78 | |
2018 | |||||||
[186] | 2018 | Patch-based CNN model | 2015 | 0.95 | – | – | |
ISLES 2015 | Stroke | 1.00 | – | – | |||
ISLES 2017 | 0.98 | – | – | ||||
[217] | 2016 | CNN | ISLES 2015 (SISS, SPES) | 0.59 on SISS, 0.77 on SPES | – | – |
Brain tumor detection using transfer learning
References | Year | Methods | Datasets/images | Results |
---|---|---|---|---|
[241] | 2019 | Pre-trained Alex-net and Google net | BRATS [2013, 2014, 2015, 2016] and ISLES-2018 | Up to 95% ACC |
[16] | 2020 | Features from VGG-19 and LBP and HOG | BRATS) [2015, 2016, and 2017] | 0.99, 1.00 and 0.99 Dice scores, respectively |
[237] | 2020 | ResNet50, InceptionV3, MobileNet -V2, NASNet and DenseNet201 | 233 MRI patient’s data | ACC of 92.9%, 92.8%, 91.8%, 99.6%, 93.1%, respectively |
[235] | 2020 | Alexnet, Resnet50, GoogLeNet, VGG-16, Resnet101, VGG-19, Inceptionv3, and InceptionResNetV2 | Harvard and local datasets | ACC of 100%, 94%, and 95.92% |
[234] | 2020 | Pre-trained visual attention model | 3064 tumor slices of local data | 95.5% ACC |
[233] | 2020 | Alex, Google and VGG | 233 MRI patient’s data | 97.3% ACC |
[232] | 2020 | VGG-19 with post-processing | 2019 BRATS series | 93.2 dice scores |
[243] | 2020 | Inception-v3 and DensNet201 | 3064 tumor slices of local data | 99.34%, and 99.51% |
[244] | 2020 | ResNet-attention gate | BRATS (2017, 2018 and 2019) | 86.5% ACC |
[247] | 2020 | VGG19 | 233 MRI patient’s data | 96.13% ACC |
[223] | 2020 | AlexNet, VGG16, ResNet18, ResNet50, VGG19, ResNet-Inception-v2, SENet, GoogleNet, ResNet101 | 233 MRI patient’s data | Up to 95% ACC |
[248] | 2020 | UNet-VGG16 | Local data | Up to 96% ACC |
[249] | 2019 | GoogLeNet | 233 MRI patient’s data | 98% ACC |
[250] | 2019 | AlexNet | Local data | 100% ACC |
[251] | 2019 | ResNet34 | Local data | 0.7380 ± 0.16 ACC |
[252] | 2020 | InceptionV3 | 233 MRI patient’s data | 99% ACC |
[253] | 2020 | InceptionV3, SqueezeNet, VGG19 and ResNet50 | Local data | Up to 97% ACC |
[254] | 2018 | AlexNet and GoogLeNet | Local data | Up to 80% ACC |
[255] | 2019 | VGGNet and ResNet | Local data | 97% ACC |
[256] | 2021 | AlexNet | 2019 BRATS | 82% AUC |
[257] | 2021 | Resnet-50, VGG-16 and Inception-V3 | Local data | ACC of 95%, 90% and 55%, respectively |
[240] | 2021 | ResNet, MobilNet-V2 and Xception | Local data | Up to 98% ACC |
[239] | 2021 | ResNet-50 model with average global pooling | Local data | Up to 97% ACC |
[238] | 2021 | 8 layers of Alex-network | Local data | 100% ACC |
Brain tumor detection using quantum machine learning
Limitations of existing’s machine/deep learning methods
Ref | Year | Methods | Brain tumor types | Results | Problems/limitations |
---|---|---|---|---|---|
[267] | 2015 | Gaussian hidden markov random field (GHMRF) | Glioma (complete, core and enhance) | 0.59 DSC on enhance tumor region | Enhance tumor region is not accurately/properly segmented using GHMRF |
[145] | 2017 | Hu moment invariants (HMI) and SVM | Glioma/non-glioma | 0.92 Specificity | The Hu moment invariants (HMI) with SVM cannot perform well on large scale dataset because SVM is employed for the recognition and SVM performs better on small scale datasets |
[268] | 2018 | Gabor filter with RF | Glioma (complete, core and enhance) | 0.84 DSC | Gabor filter with RF method fails in detecting tumor with small volume and its boundaries |
[269] | 2019 | ALI, lesionGnb and LINDA methods | Stroke | 0.50 Dice score (DC) | Due to variable size and stroke territory, ALI, lesionGnb and LINDA methods fail to detect the small lesion region |
[220] | 2019 | DeepMedic and U-Net | Glioma (complete, core and enhance) | 990.64 DSC | The enhance tumor region is segmented with less dice score compared to complete and non-enhance region |
[266] | 2020 | Progressive growing of generative adversarial networks (GANs) (PGGANs) | Glioma/non-glioma | 0.91 Accuracy | PGGANs method fails to differentiate tumor/non-tumor features in 25% cases |
[270] | 2020 | Unsupervised probabilistic model | Glioma and stroke | 0.34 ± 0.20 AUC | Fail to detect the small and unobvious brain lesions |
[271] | 2020 | 3D Squeeze-and-Excitation V-Net | Glioma (complete, core and enhance) | 0.74 DSC enhance tumor,0.89 DSC complete tumor,0.80 DSC non-enhance | 3D Squeeze-and-Excitation V-Net does not provide accurate tumor prediction in some cases |
[272] | 2021 | Adaptive deep fuzzy neural model | Glioma/non- glioma | 0.99 ACC | Computationally exhaustive still need to work on light weight model for accurate brain tumor detection |
Research findings and discussion
-
The size of a brain tumor grows rapidly. Therefore, tumor diagnosis at an initial stage is an exigent task.
-
Brain tumor segmentation is difficult owing to the following factors.
-
MRI image owing to magnetic field fluctuations in the coil.
-
Gliomas are infiltrative, owing to fuzzy borders. Thus, they become more difficult to segment [43].
-
Stroke lesion segmentation is a very intricate task, as stroke lesions appear in complex shapes and with ambiguous boundaries and intensity variations.
-
The optimized and best feature extraction and selection is another difficult process inaccurate classification of brain tumors.