There are many challenging cases, in which the single segmentation is not enough. The paranasal sinuses and the congenital heart defects are notable examples. My usual workflow was to segment whatever I can as good as it's possible, to clean the unnecessary structures and the artefacts, to export the segmentation as stl 3d model and then to "CAD my way around".  This is solid philosophy for simple, uncomplicated models, but for complex structures with a lot of small details and requirement from the client for the highest quality possible, this is just not good enough, especially for a professional anatomist like myself. Then I started to exploit the simple fact, that you're actually able to export the model as stl, to model it with your CAD software and then to reimport it back and convert it into label map again. I called this "back and forth technique". You can model the finest details on your model and then you can continue the segmentation right where you need it, catching even the slightest details of the morphology of the targeted structure. This technique, combined with my expertise, gives me the ability to produce the best possible details on some of the most challenging cases, including nasal cavity, heart valves, brain models etc. etc.
To use this technique, just import the stl file, convert it into a label map (for 3D slicer -  segmentation module/ export/import models and label maps). 

The main advantages of this technique are:
1. You can combine the segmentation with the most advanced CAD functions of your favorite software. Two highly specialized programs are better than one "Jack of all trades" (cough cough Mimics cough cough)
2. Advanced artefact removing.
3. Advanced small detail segmentation and modelling.
4. Combined with several markers (separate segmentations, several voxels in size) on the nearby anthropometric points, this technique increases the accuracy of the final product significantly. Without points of origin, the geometry of your model will go to hell, if you're not especially careful (yes, I'm talking about the 3D brushes in Slicer).
5. You can easily compare the label map with the 3d model, converted back. Every deviation, produced during the CAD operations will be visible like a big, shining dot, which you can easily see and correct. This is one of the strongest quality control techniques.
6. You can create advanced masks with all the geometrical forms you can possibly imagine, which you can use for advanced detail segmentation. Those masks will be linked with the spatial coordinates of the targeted structures - the stl file preserves the exact coordinates of every voxel, which was segmented.
7. You can go back and forth multiple times, as many as you like.
8. This technique is more powerful than the best AI, developed by now. It combines the best from the digital technologies with the prowess of the human visual cortex (the best video card up to date).

The main disadvantages are:
1. It's time consuming.
2. It produces A LOT of junk files.
3. Advanced expertise is needed for this technique. This is not some "prank modelling", but an actual morphological work. A specialized education and practical experience in the human anatomy, pathology and radiology will give you the best results, which this technique can offer.
 4. You need highly developed visual cortex for this technique (dominant visual sense). This technique is not for the linguistic, spatial-motor, olphactory etc. types of brains. Recent studies confirms, that a part of the population have genetically determined bigger, more advanced visual cortex (The human connectome project, Prof. David Van Essen, Washington University in Saint Louis). Such individuals become really successful cinematographers, designers, photographers and medical imaging specialists. The same is true for all the other senses, but right now we're talking about visual modality and 3D intellect (I'm sorry, dear linguists, musicians, craftsmen and tasters). It's not a coincidence that I have so many visual artists in my family (which makes me the medical black sheep). But if you don't have this kind of brain, you can still use the technique for quality control and precise mask generation. Just let the treshould module or the AI to do the job for you in the coordinates, in which you want (You should really start using the Segment Editor module in Slicer 3D).
5. You really need to love your work, if you're using this technique. For the usual 3D modelling you don't need so many details in your model and to "CAD your way around" is enough for the task.
6. You should use only stl files. For some reason, the obj format can't preserve the spatial geometry as good as the stl format. Maybe because the stl is just a simple map of vertex coordinates and the obj contains much more sophisticated data. The simple, the better.

On the picture - comparison of the semilunar valves, made by treshould segmentation at 250-450 Hounsfield units (in green) and modelled and reimported model (in red).