Table 2.

Comparison between different cIMT regional segmentation methods for given attributes

Attributes	Conventional methods (1.1)	Machine learning (1.2.1)		Deep learning (1.2.2)
Types	Active contour, level sets, edge detection, dynamic programming (1.1.*)	Supervised (1.2.1.1)	Geometric: ANN, SVM (1.2.1.1.1)	Supervised (1.2.2.1)	CNN, FCN (1.2.2.1.*)
			Logical: decision trees (1.2.1.1.2)
			Probabilistic: Bayesian (1.2.1.1.3)	Unsupervised (1.2.1.2)	Auto-encoders, deep belief nets (1.2.2.2.*)
		Unsupervised (1.2.1.2)	K-means clustering (1.2.1.2.1)
Description	Techniques employing imaging features and processing for the cIMT regional segmentation	The two-stage intelligence-based paradigm in which image features are extracted and then used by the ML model for training and testing for the cIMT region identification		Mimics human visual cortex with numerous hierarchical network of neurons extracting high-level features directly from images for recognition of the cIMT region
Advantages	1. Fastest among all methods for region estimation as no training and testing 2. Simple	1. Uses intelligence in the form of learning from experience/training and applies to unknown instances 2. Somewhat generalized and can be applied to instances within the similar domain 3. Faster than deep learning		1. Independent of feature extraction algorithms 2. Can be scaled up to recognize/characterize millions of images 3. Generalized and can be used to multiple domains
Disadvantages	1. Task-specific and cannot be generalized 2. Low accuracy	1. Dependent upon the quality of features for better accuracy 2. The accuracy curve diminishes when scaled up to a large number of instances		1. Costly in terms of computation time and memory 2. Easily overfits and therefore requires different techniques for prevention of overfitting