Going from a CT series to a model you can rotate, measure, or print is a conversion job. You isolate anatomy, build a mesh, clean it, and export a format the next tool accepts. This is not the same as reconstructing slices from projections. That already happened on the scanner.
In the mid-1980s, stacking 50–60 CT slices into a 3D view could take more than 24 hours and produce on the order of 84 frames. Hardware like the Pixar Image Computer made it possible, not fast. CTisus has that history. Today the same path runs in minutes if the scan was acquired for 3D in the first place.
The conversion path
- Acquire for the model. Thin slices, little or no gap, isotropic voxels if you can get them. Thick survey slices make stair-steps. For cardiac work, ECG gating matters: ungated motion wrecks small structures (coronaries). Voxel isotropy and motion are the two usual quality killers; see this conversion review.
- Segment. Pick the tissue. The basic tool is a Hounsfield threshold (bone is bright, air is black). Automatic algorithms are fast and often wrong at thin cortex, contrast edges, and metal. Semi-automatic (you seed, the tool grows) is the usual clinical compromise. Manual slice-by-slice is slow and still the fallback for ugly cases.
- Mesh. Marching cubes (or a sibling) turns the binary volume into a triangle surface. That mesh is what CAD, Blender, and a printer speak.
- Repair and export. Close holes, remove islands, decimate if the file is huge. STL for print. OBJ if you still need texture or a DCC. Check watertight before you send it to a printer.
Threshold choice is not cosmetic. A comparison of alignment methods reported mean discrepancies from 0.004 mm to 0.141 mm depending on the threshold/method pair (PMC3047291). For a guide that sits on bone, that range is the difference between “fits” and “doesn’t.”
Software
| Software | What it is good at | Learning curve | Price | Best for |
|---|---|---|---|---|
| 3D Slicer | Full research pipeline: DICOM in, segment, visualize, some modeling | Steep | Free | Research, teaching, people who will read docs |
| Materialise Mimics | Medical segmentation plus CAD export, device-oriented modules | Steep | High | Implants, guides, regulated device work |
| ITK-SNAP | Focused segmenter, simple 3D view | Moderate | Free | Getting a mask out of a volume |
| Blender | Mesh cleanup, sculpt, render. Not a DICOM workstation | Steep | Free | After someone else already made an STL/OBJ |
| Meshmixer | Hole fill, island remove, hollowing, print prep | Moderate | Free | Making a mesh survive a printer |
Pick the tool from the output, not the brand. Need a 510(k)-adjacent implant workflow? Mimics (or a similar medical CAD stack). Need a mask for a paper? Slicer or ITK-SNAP. Need a pretty still or a cleaned print file? Export the mesh and finish in Blender or Meshmixer. PYCAD sits on the other side of this: a custom web viewer and case workflow when the 3D view has to live next to PACS and reports, not on one engineer’s desktop. Case studies.
What the model is for
- Surgical planning and rehearsal. Ortho reconstructions, maxillofacial trauma, some neuro cases. People print the bone and hold it.
- Patient-specific hardware. Cranial plates, osteotomy guides, some joint components. The mesh is the CAD input.
- Teaching and consent. A physical or on-screen model of this patient’s anatomy, not a textbook drawing.
AI helps on step 2 (segmentation) and sometimes on sparse-view reconstruction of the volume itself. It does not remove the need to check the mesh against the source slices. Training data and validation are still the limit: a network that never saw your implant or your scanner protocol will miss cortex.
Photo-based 2D-to-3D (one snapshot, photogrammetry, Blender) is a different pipeline. That is how to turn a 2D image into a 3D model. MRI volumes follow the same mesh steps once you have a segmentable series: 3D MRI reconstruction.