Table 4 Pros and cons of DL and TA in 3D reconstruction of medical images
Model | Type | Objective | Pros | Cons |
|---|---|---|---|---|
Active Contour [39] | TA | Segmentation and contour detection | Reset and re-evaluation is fast, dynamic and verifiable with a GUI | Requires initial seed region, for correct segmentation more than one trials of seeds at proper location is needed, human supervision is needed |
U-Net [61] | DL | Segmentation and contour detection | Does not require initial seed region, highly accurate and much superior to TA, automatic feature detection, available as pre-trained model | Accuracy is dependent on training examples, incurs huge computation cost, addition of new data is difficult and requires fine tuning, initial target masks are required |
Marching Cube [52] | TA | 3Dmesh generation | Easy to implement, tailor made for medical image reconstruction, requires very little storage of memory | Generated output is rough and not smooth, higher volume of CT slices generates higher level of details, requires complete set of segmented volume |
DNF-Net [149] | DL | 3Dmesh generation | Filters noisy polygon normal, accurately predict orientation and direction of surface normal, superior to TA | Insufficient training examples results in unexpected results, does not generate mesh by itself, requires update algorithm to convert normal into mesh |
B-Spline surfaces [45] | TA | 2D-3D registration | Easy to implement, requires no prior training examples | Requires equal number of control points from every CT scan within a volume, human intervention is needed |
Voxelmorph [145] | DL | 2D-3D registration | Automatic and unsupervised | As an atlas based approach it is reference dependent |