Table 1.

The table reports the main characteristics and findings of the studies focused on lesion detection and the differential diagnosis of brain tumors.

Author and Year	Country	N. Patients	Database	MRI Sequences and Clinical Data	AI Model	Task	Main Results	Main Limitations
Park et al. [46]	South Korea	188 (917 lesions)	Institutional brain MRI database	3D-GRE, 3D-BB	DL model based on 3D U-net	BM detection (3D-BB + 3D-GRE vs. 3D-GRE)	3D-BB + 3D-GRE model sensitivity = 93.1% 3D-GRE model sensitivity = 76.8%, (p < 0.001)	Single-center, retrospective study, small data size, 3D-BB sequences may have limited availability in MRI scanners, model mostly trained on patients with metastases
Swinburne et al. [50]	USA	26	Institutional brain MRI database	DWI, DSC, DCE	MLP (Multilayer Perceptorn) model using VpNET2	GBM vs. BM vs. PCNSL	Increase in 19.2% in correct diagnoses in cases where neuroradiologists disagreed	Manual tumor segmentation, sample size, no evaluation with an independent test cohort
Skogen et al. [52]	Norway	43	Institutional brain MRI database	DTI (FA and ADC)	Commercially available texture analysis research software (TexRAD)	GBM vs. BM	The heterogeneity of the peritumoral edema was significantly higher in GBMs (sensitivity 80% and specificity 90%)	Retrospective study, analysis of a single slice, the manual drawn of the ROI
Han et al. [53]	China	350	Institutional brain MRI database (two centers)	T1C, clinical data (age, sex), routine radiological indices (tumor size, edema ratio, location)	AI-driven model using logistic regression model	GBM vs. BM (lung cancer and other sites)	Combination models superior to clinical or radiological models (AUC: 0.764 for differentiation and 0.759 for differentiation between MET-lung and MET-other in internal validation cohorts)	Radiomic only based on T1-enhanced images, retrospective study, many small groups of metastases from other than lungs
Ortiz-Ramón et al. [55]	Spain	67	Institutional brain MRI database	IR-T1	RF model	Differentiate the primary site of origin of brain metastases	Images quantized with 32 gray-levels (AUC = 0.873 ± 0.064). differentiating lung cancer from breast cancer (AUC = 0.963 ± 0.054) and melanoma (AUC = 0.936 ± 0.070)	Small set of BM, single-center study,
Stadlbauer et al. [59]	Austria	167	Institutional brain MRI database	Standard MRI (FLAIR, T1C), advanced MRI (DWI, DSC), physiological MRI (VAM = vascular architecture mapping)	Nine commonly use ML (SVM, DT, kNN, MLP, AdaBoost, RF, bagging)	GBM vs. HHG (anaplastic glioma) vs. meningioma vs. PCNSL vs. BM	Adaptive boosting and random forest + advanced MRI and physiological MRI data were superior to human reading in accuracy (0.875 vs. 0.850), precision (0.862 vs. 0.798), F-score (0.774 vs. 0.740), AUROC (0.886 vs. 0.813), and classification error (5 vs. 6)	Small sample size, single MRI scanner and traditional ML
Ucuzal et al. [60]	Turkey	233	Open-source dataset from https://figshare.com (accessed on 01 January 2022).	T1C	CNN from DL algorithm, developed web-based software (Python programming language and TensorFlow, Keras, Scikit-learn, OpenCV, Pandas, NumPy, MatPlotLib, and Flask libraries)	Glioma vs. Meningioma vs. Pituitary lesions	All the calculated performance metrics are higher than 98% for classifying the types of brain tumors on the training dataset	Small size, not healthy individuals, the selection and creation of these algorithms may require a lot of time and experience
Pavabvash et al. [65]	USA	256	Institutional brain MRI database	T1, DWI, T2, FLAIR, SWI, DSC, T1C	Naïve Bayes, RF, SVM, CNN	Differentiation of posterior fossa lesions (Hemangioblastoma, Pilocytic Astrocytoma, Ependymoma, Medulloblastoma	The decision tree model achieved greater AUC for differentiation of pilocytic astrocytoma (p = 0.020); and ATRT (p = 0.001) from other types of neoplasms	Small number of rare tumor types, lack of molecular subtyping in medulloblastoma and ependymoma, manual segmentation, acquisition in different field strengths
Verma et al. [67]	Switzerland	32	Institutional brain MRI database	DSC, T1CI	DTPA-method with different texture parameters	GBM vs. PCNSL, tumefactive multiple sclerosis	The texture parameters of the original DSCE-image for mean, standard deviation and variance showed the most significant differences (p-value between <0.00 and 0.05) between pathologies	Small size, smaller TOI in MS, manual segmentation
Han et al. [68]	China	57	Institutional brain MRI database	T1, T2	t-test and statistical regression (LASSO algorithm) to develop three radiomic models base on T1 WI, T2 WI and a combination	LGG vs. multiple sclerosis	T2WI and combination models achieved better diagnostic efficacy, with AUC of 0.980, 0.988 in primary cohort and that of 0.950, 0.925 in validation cohort,	Retrospective study, small size, single scanner, unknown etiology of inflammation
Qian et al. [69]	China	412	Cancer Genome Atlas (TCGA); retrospective dataset from Beijing Tiantan Hospital	T1C	Radiomic features extraction, ML	GBM vs. single BM	SVM + LASSO classifiers had the highest prediction efficacy (AUC, 0.90)	Retrospective study; imaging data from multiple MRI systems; only CE sequences were used
Bae et al. [70]	Korea	166 (training) + 82 (validation)	retrospective institutional brain MRI database	T2, T1C	DL using radiomic features	GBM vs. single BM	DNN showed high diagnostic performance, with an AUC, sen, spec, and acc of 0.956, 90.6%, 88.0% and 89.0%	Automated tumor segmentation, not included advanced sequences, heterogeneous MR scanner types
Adu et al. [61]	China		Brain Tumor Dataset. Figshare (3064 images)	T1C	CapsNets (dilated capsulenet)	Detection + classification	Acc.: 95%	Not enough comparisons and experiments with confusion matrix
Abiwinanda et al. [43]	Indonesia		Brain Tumor Dataset. Figshare (3064 images)	T1C	CNN	Classify into three types	Acc.: 98%	Complexity of pre-processing