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

Table 4

Segmentation techniques: DeepLab V3, SegNet, and Fast Convolutional Network (FCN).

Method	Advantages	Disadvantages	Dataset	DC	ID	Modality	App	Ref
DeepLab V3 [41, 42]	(i) Allows us to enlarge the fields of view of filters to incorporate large context; (ii) Preserves spatial information	(i) Has no postprocessing step conditional random fields; (ii) Does not scale well for large or deeper layers if GPU memory is limited	CHAOS	81.0	3D	CT and MRI	Kidney	Guo et al.
SegNet [43]	(i) Low memory requirement during both training and testing; (ii) Improved boundary delineation; (iii) Reduced number of parameters enabling end to end training	(i) Both input image and output segmentation have fixed resolution	OASIS	91.47	3D	MRI	Brain	Khagi et al.
FCN [44, 45]	(i) Ability to make predictions on arbitrarily sized inputs; (ii) End to end trainable fast and improved performance	(i) Direct predictions are typically in low resolution resulting in fuzzy object boundaries; (ii) Suitable mainly for object detection, not object classification (used for local rather than global tasks)	DRIVE	95.33	3D	Funduscopy	Retinal vessels	Cai et al.