Table 2.

DCNN-based segmentation models brief description (sorted by year published//DCNN performance).

Ref	Year	Tumor Type	Task	Model Name	Image Type	Model Desc.	Software	Hardware	Dataset	Instances	Performance
[70] **	2020	Glioma, Meningioma, Pituitary	Segmentation	ELM-LRF	MRI	Implemented an enhanced softmax loss function that is more suitable for multiclass applications.	Python 3.6; Keras	–2.8 GHz Intel Core i7 7th gen processor with 16 GB RAM and 4 GB –NVIDIA 1050 memory	Brain Tumor Dataset [58]	3064	99.54%, 98.14%, 98.67% (Per Tumor Type)
[69]	2020	Glioma, Meningioma, Pituitary	Segmentation	ResNet50	MRI	Glioma, meningioma, and pituitary tumor segmentation with the ResNet50 architecture.	–Python 3.6; Keras 2.2.4; –Tensorflow 1.13	–NVIDIA GeForce RTX 2070 –GPU; Intel i5-9600K @ 3.7 GHz and 16 GB RAM	Brain Tumor Dataset [58]	3064	99%
[73]	2020	Glioma	Segmentation	HCNN; CRF-RRNN	MRI	The composite architecture of HCNN to capture mixed scale context and CRF-RRNN reconstruct a global segmentation.	N/A	N/A	–BraTS2013 [59] –BraTS2015 [64]	220 HGG; 50 LGG	98.6%
[74]	2019	Glioma	Segmentation	FCNN; DMD	MRI	Enhanced FCNN with batch normalization and DMD features to provide spatial consistency. Fisher vector encoding method for texture invariance to scale and rotation.	Caffe	–CPU Intel Core i7 –3.5GHz, GPU NVIDIA GeForce GTX1070	BraTS2015 [64]	220 HGG; 50 LGG	91%
[86] **	2018	Glioma	Segmentation	P-Net; PC-Net	MRI	Addresses zero-shot learning by taking user input bounding boxes and scribbles to fine-tune segmentations.	Caffe	–2 8-core E5-2623v3 Intel –Haswell, a K80 NVIDIA GPU and 128GB memory	BraTS2015 [64]	220 HGG; 50 LGG	86.29%
[75]	2019	Glioma	Segmentation	FCNN	MRI	A novel N3T-spline utilizes is used to preprocess 3D input images. GLCM extracts feature vectors and are inputs into a CNN.	MATLAB R2017a	N/A	BraTS2015 [64]	220 HGG; 50 LGG	N/A
[71]	2020	N/A	Segmentation	3-layer DCNN *	MRI	Utilized Otsu thresholding to create a novel skull stripping algorithm. GLCM and a three-layer CNN segments the stripped images.	MATLAB R2018b	N/A	IBSR [89]	18	98%
[72]	2020	Glioma	Segmentation	Automatic Detection and Segmentation of Tumor (ADST) *	3D MRI	Region Growing and Local Binary Pattern (LBP) operators are used to build a feature vector that is then segmented	N/A	N/A	BraTS2018 [61]	210 HGG; 75 LGG	87.20% for HGG; 83.77 for LGG (Average Jaccard)
[81] **	2019	Glioma	Segmentation	Hourglass Net	MRI	Enhanced Hourglass Network with added residual blocks and novel concatenation layers.	N/A	NVIDIA TITAN X GPU	BraTS2018 [61]	210 HGG; 75 LGG	92%
[83] **	2019	Glioma	Segmentation	XGBoost; U-Net; DAU-Net (Domain Adaptive U-Net) *	MRI	Implementation of a U-Net variation using instance normalization to boost domain adaptation.	PyTorch	4 NVIDIA Titan Xp GPU cards	BraTS2018 [61]	210 HGG; 75 LGG	91% (Whole Tumor)
[82]	2019	Glioma	Segmentation	MC-Net; OM-Net	MRI	Ensemble network of several MC-Net and OM-Net variations. Attention mechanisms are added to increases sensitivity to relevant channel-wise interdependencies	N/A	N/A	BraTS2018 [61]	210 HGG; 75 LGG	90% (Whole Tumor)
[84] **	2019	Glioma	Segmentation	U-Net	MRI	The U-Net variation that uses batch normalization and residual blocks to improve performance on neurological images.	N/A	–Intel Xeon E5-2650 CPU@ 2.00 GHz (64 GB) and NVIDIA Quadro –4000–448 Core CUDA (2 GB) GPU	BraTS2018 [61]	210 HGG; 75 LGG	86.8% (Whole Tumor)
[85]	2019	Glioma	Segmentation	U-Net	MRI	Extension of the U-Net to train with “mixed supervision”, meaning both pixel-wise & image-level ground truths to achieve superior performance.	N/A	N/A	BraTS2018 [61]	210 HGG; 75 LGG	N/A for the entire dataset
[77]	2019	Glioma, Meningioma, Pituitary, and Negative	Classification	–CNN, –Xception, VGG16, VGG19	MRI	Several of the pre-trained models (simple CNNs, Xception, VGG16, and VGG19) were fused together in a composite architecture.	Keras	N/A	N/A	1167	98.89%
[87] **	2020	TBI (Traumatic Brain Injury)	Segmentation	CNN	CT	3D CNN architecture to create voxel-wise segmentation of TBI CT scans.	N/A	N/A	CENTER-TBI (Datasets 1 & 2) [90]; CQ500 [91]	539; 500 (Patients)	94%
[78]	2019	N/A	Segmentation	M-SVM; CNN	MRI	SGLDM and M-SVM are applied to extract and classify MRI scans. CNN is then applied to segment the extracted feature vectors.	N/A	N/A	N/A	40	84%
[76]	2019	N/A	Segmentation	SWT; GCNN	MRI	Dataset is preprocessed with a novel skull stripping. Features are extracted with SWT, classified with a Random Forest implementation and finally segmented with GCNN.	N/A	N/A	BRAINIX [92]	2457	98.6% (SSIM Score)
[80] **	2018	Glioma	Segmentation	U-Net	MRI	HPU-Net enhances the traditional U-Net with multiscale images and image pyramids.	Keras; Tensorflow	NVIDIA Titan X GPU	–BraTS2015 [64] –BraTS2017 [93]	430 HGG; 145 LGG	71% and 80% (Respective to dataset)
[88] **	2015	Glioma	Segmentation	ImageNet LSVRC 2013	H&E Histology	Patches are extracted from large histopathology scans and passed into ImageNet LSVRC 2013 architecture. Linear SVM classifier pools extracted feature vectors.	N/A	N/A	MICCAI 2014 [94]	35	84%

* Model names with asterisks are not defined in the original papers and names were assigned based on the models applied. Note: For abbreviations description in this table please refer to the list of abbreviations on the back part of this article (before References). ** the references with “**” mean that the results achieved by their methods or the dataset used have been validated/supervised by specialists (e.g., pathologists/radiologists).