Table 1.

Comparison between Vess-Net and previous methods for vessel segmentation.

Type	Methods	Strength	Weakness
Using handcrafted local features	Vessel segmentation using thresholding [23,24,28,29,31,33,34,35,36]	Simple method to approximate vessel pixels	False points detected when vessel pixel values are closer to background
	Fuzzy-based segmentation [25]	Performs well with uniform pixel values	Intensive pre-processing is required to intensify blood vessels’ response
	Active contours [26,30]	Better approximation for detection of real boundaries	Iterative and time-consuming processes are required
	Vessel tubular properties-based method [32]	Good estimation of vessel-like structures	Limited by pixel discontinuities
	Line detection-based method [27]	Removing background helps reduce false skin-like pixels
Using features based on machine learning or deep learning	Random forest classifier-based method [37]	Lighter method to classify pixels	Various transformations needed before classification to form features
	Patch-based CNN [38,42]	Better classification	Training and testing require long processing time
	SVM-based method [41]	Lower training time	Use of pre-processing schemes with several images to produce feature vector
	Extreme machine-learning [39]	Machine learning with many discriminative features	Morphology and other conventional approaches are needed to produce discriminative features
	Mahalanobis distance classifier [40]	Simple procedure for training	Pre-processing overhead is still required to compute relevant features
	U-Net-based CNN for semantic segmentation [43]	U-Net structure preserves the boundaries well	Gray scale pre-processing is required
	Multi-scale CNN [44,47]	Better learning due to multi-receptive fields	Tiny vessels not detected in certain cases
	CNN with CRFs [45]	CNN with few layers provides faster segmentation	CRFs are computationally complex
	SegNet-inspired method [46]	Encoder and decoder architecture provides a uniform structure of network layers	Use of PCA to prepare data for training
	CNN with visual codebook [48]	10-layer CNN for correlation with ground truth representation	No end-to-end system for training and testing
	CNN with quantization and pruning [49]	Pruned convolutions increase the efficiency of the network	Fully connected layers increase the number of trainable parameters
	Three-stage CNN-based deep-learning method [50]	Fusion of multi-feature image provides powerful representation	Usage of three CNNs requires more computational power and cost
	Modified U-Net with dice loss [51]	Dice loss provides good results with unbalanced classes	Use of PCA to prepare data for training
	Deformable U-Net-based method [52]	Deformable networks can adequately accommodate geometric variations of data	Patch-based training and testing is time-consuming
	PixelBNN [53]	Pixel CNN is famous for predicting pixels with spatial dimensions	Use of CLAHE for pre-processing
	Dense U-Net-based method [54]	Dense block is good for alleviating vanishing gradient problem	Patch-based training and testing is time-consuming
	Cross-connected CNN (CcNet) [55]	Cross-connections of layers empower features	Complex architecture with pre-processing
	Vess-Net (this work)	Robust segmentation with fewer layers	Augmented data necessary to fully train network