Table 1.

Literature review.

Sr. No.	Methodology	Results	Future Directions
1	Three layered feed forward ANNs and two real world problems are set as a benchmark to access the performance of Group Search Optimizer (GSO) [27].	GSOANN has a far better performance as compared to regular ANN.	—–
2	A hybrid model of DSA and DL to help improve the relationship of computer science and bioinformatics [28].	Differential Search Algorithm (DSA) and DL can help produce more xylitol for sugar free gums.	Computational biologists and computer scientist can together produce a hybrid model using deep learning OA.
3	In auto-encoders like VVG-9 and CIFAR-10, they design some experiments to study the properties of RMSProp and Adam against Nesterov’s Accelerated Gradient method [29].	On very high values of $\beta$1 = 0.99 Adam outperforms lower training and test losses, whereas with $\beta$1 = 0.9, NAG performs better.	Advance theory in getting more better results by getting $\beta$1 close to 1.
4	Different optimization algorithms are studied by side CNN architecture [30].	Among 7 optimizers, on the LeNet architecture, Adam provides the smallest MSE whereas SGD and Adagrad failed.	Can build analytical protable image devices
5	Constructed a few illustrative binary classification problems and examined empirical generalization capability of adptive methods agaisnt GD.	Solutions found by adaptive methods generalize worse than GSD.	Adaptive methods should be reconsidered.
6	Energy Index based Optimization Method (EIOM) that automatically adjusts the learning rate in backpropagation [31].	EIOM proves to be the best when compared with state-of-the-art optimzation methods.	—–
7	A non-asymptotic analysis of the convergence of two algorithms: SGD and simple averaging [32].	The analysis suggests that the learning rate is proportional to the inverse of the number of iterations.	Differential and non-differential stochastic
8	Adaptive learning rate and laplacian approach have been proposed for Deep Learning in MLP [33].	Improved classification accuracy	—–
9	Proposed a fundamental approach for anatomical, celluler stuctures, and tissue segmentation using CNN through image patches measuring 13 × 13 voxels [34].	On different data sets, comparing the six commonly used tools (i.e., ROBEX, HWA, BET, BEaST, BSE, and 3dSkullStrip), they achived the highest average specifity.	Can be performed on most advanced tools and used a real time data set to get better result.
10	Used a pretrained CNN model on augmented and orginal data for brain tumor classification [35]	They achieved 90.67 accuracy before and after data augmentation on the proposed methed and compared with most advanced methods	Used light weight CNN to entend their work for fine-grained classification differential stochastic.
11	A CapsNet for brain tumor classification and investigation of the overfitting problem based on CapNet [36].	On 10 epochs, they achieved 86.56% accuracy, with the comparative analysis with CNN learning rate proportional to the inverse of the number of iterations.	In the future, investigations on the effects of more layers on the classification accuracy will be performed.
12	A review on deep learning techniques in the field of medical images classification [37]	They discussed in detail the deep learning approaches and their suitability for medical images. The learning rate is proportional to the inverse of the number of iterations.	Further research is required to apply the techniques to the modalities, where these are not applied.
13	GA-SVM and PSO-SVM method used to classify heart disease [38].	GA and particle swarm optimization (PSO) algorithms combined with SVM achieved a high accuracy.	—–
14	Applied U-NET approach using BraTS2017 data set and prediction of patient survival [39]	89.6% Accuracy achieved with less computational time	—–
15	Two-way path architecture based on CNN for brain tumor segmentation on the BraTS 2013 and 2015 data sets [3]	Input cascaded CNN got a high accuracy with 88.2% on the comparitive analysis with other architechtures.	Further improved the results with increasing architechture layers and data set.