Automated Techniques for the Interpretation of Fetal Abnormalities: A Review

Table 2

Comparative analysis of important fetal image segmentation techniques.

Segmentation techniques	Advantage	Limitation	References
Constrained probabilistic boosting tree	The results are based on the tree structure, so segmented biometric parameters are measured accurately.	The process of multistage decision and the data input is in binary form.	[9, 34, 35]
Fuzzy decision system	The detection is based on fuzzy boundary, and all parameters are boundary sensitive.	The fuzzy system is based on a series of If-Then rules, making the system complicated.	[38]
Class-separable sensitive approach	The fetal biometric parameter shape is of different types so the class-separable approach is good.	US image is having some noise, and it is very much sensitive to noise.	[39, 40]
Thresholding-based morphological operator	Advantage of thresholding lies in its simplicity, which involves minimal implementation and computational requirements.	It is sensitive to noise, and it cannot be an effective segmentation technique for US medical images.	[14, 66, 67]
Edge detection algorithm	The amount of data to be processed is reduced, and the analysis of images is simple. Besides, at the same time it preserves useful information about object boundaries.	The masks used by different operators act as a high-pass filter, which tend to amplify the noise.	[67, 68, 77, 78]
Active contour model	It can generate the closed parametric curve directly from the images by calculating the external force. It also includes the robustness against the noise (internal force).	The initial contour is placed manually, so the method is sensitive. Problems are associated with initialization of contour and convergence to their boundary concavities.	[52–54, 14, 63]
Wavelet transform	This approach is based on the texture of the object so the results are accurate.	The fetal parameters is of various sizes so sometimes the discrimination is emblematic.	[51]
Graph-based approaches	This approach is good because the whole image is considered and the evaluation of parameters is closer to the expert results.	In this approach, few clicks are placed manually for continuous min-cut partition of the graph.	[13, 56–58]
Neural network	The NN can be applied to any classification/recognition problem by modifying only the training set. So easily the network can be trained.	There are various types of classifier used in NN, so the selection of a proper algorithm and classifier gives good results.	[77–79, 83, 84]
Level set	All level sets yield a nice representation of regions, without the need of a complex data structure.	A level set function is restricted to the separation of two regions. As soon as two regions are considered, the level set idea loses part of its attractiveness. Results vary due to initial contour placement.	[81, 82]