Pancreatic Cancer Survival Prediction: A Survey of the State-of-the-Art

Table 3

Segmentation techniques: dice coefficient (DC), image dimension (ID), application (App), reference (Ref), convolutional neural network- (CNN-) conditional random field (CRF), named entity recognition (NER), maximum entropy Markov models (MEMM), and hidden Markov model (HMM).

Method	Advantages	Disadvantages	Dataset	DC	ID	Modality	App	Ref
CNN-CRF [36, 37]	(i) Flexible enough in terms of feature selection; (ii) Better for NER than MEMM and HMM	(i) High computational complexity of the algorithm training stage algorithm; (ii) Difficulty in retraining the model when new training data is available	Data in hospitals	86.0	2D	MRI	Brain tumour	Feng et al.
U-Net [33, 38]	(i) Provides pixel-accurate semantic segmentation; (ii) It is fast to compute; (iii) Its architecture is easy to understand	(i) Significant memory requirement as lower level features have to be stored for further concatenation in the upsampling phase	1245 CT images	73.6	3D	CT	Pulmonary nodules	Tong et al., Du et al.
E-Net [39]	(i) Significantly faster; (ii) Provides a high frame rate for real-time applications; (iii) Small storage requirements alleviating the need for model compression	(i) The use of convolutional layer factorization increases the number of kernel calls making each of them smaller	Achieva scanner (Philips Healthcare, Best, The Netherlands) with a pelvic phased-array coil (8 channel HD Torso XL)	90.9	3D	MRI	Prostate	Comelli et al.
V-Net [40]	(i) Fully CNN and suitable for volumetric medical image segmentation; (ii) The residual function is learnt	(i) Location information is lost in the compression path	PROMISE 2012 Challenge	82.39	3D	MRI	Prostate	Milletary et al.