Guide · Settings
Settings
Open Settings in the sidebar to choose which models read the film and how processing runs. The defaults suit most films. The settings worth changing are Crop localizer for lumbar-only films, Low memory on smaller machines, and the processor on Windows machines with a graphics card.
At a glance
| Setting | Default | Change it when |
|---|---|---|
| Theme | Light | You prefer a dark interface. |
| L1–L5 model | U-Net | You want to compare against HRNet on lumbar films. |
| Processing mode | Standard | The computer is short on memory. |
| Processor | CPU | On Windows, a DirectX 12 graphics card is available. On a Mac, CPU is the only option. |
| Crop localizer | On | Every film is lumbar-only and already framed on the spine. |
| Crop method | Crop search | You want to try the trained crop model. |
| Toolbar removal | Off | Films are PACS screenshots with a toolbar along the bottom. |
- Settings are locked while anything is processing. Change them before a run or batch, or after it finishes.
- Changes apply to the next run. Results already saved are not recalculated. Run a study again to see the effect.
- The Processing settings are saved and persist across restarts. Theme and the L1–L5 model reset to their defaults each time the app opens.
Models
The Models block shows which model reads each structure. Only the lumbar vertebrae have a choice.
L1–L5 · U-Net or HRNet
U-Net (the default) outlines each vertebral body and reads its corners off the outline. HRNet places each corner directly. With HRNet, the U-Net outline still checks which levels are actually on the film, so levels outside a cropped film are left out rather than guessed.
Fixed models
The femoral heads use a U-Net and the S1 endplate a Keypoint R-CNN, with no alternative. Cervical and full-spine films use their own HRNet models whatever the L1–L5 choice.
Each study's Analysis header names the model that produced its numbers, so a library measured with both models can still be told apart.
Processing mode
Standard · default
Keeps every model loaded, so repeated films and batches run faster. Uses up to four CPU threads.
Low memory
Loads one model at a time and unloads it before the next. Choose 1 to 4 CPU threads (2 by default). It can be slower, but needs much less memory at once.
Both modes use the same models at the same resolution and gave identical measurements in testing. Low memory lowers the peak; it does not guarantee a fixed limit, and one model must still fit in memory.
Processor
CPU is the default and the reference path the validation figures were measured on. On Windows, any DirectX 12 graphics card SpineContour can use is listed as GPU · its name. No CUDA install is needed. macOS, including Apple Silicon, runs on the CPU only.
- Choose the GPU here, not in Windows. Windows' and NVIDIA's per-program graphics settings control drawing on screen, never the models.
- The first GPU run is slow. Before using a card, SpineContour checks that all six models give the same results on it as on the CPU. This can take several minutes, and runs again after a driver update.
- A failure falls back to the CPU. If the card fails the check or errors during a run, the whole film is processed again on the CPU and a message says so. A result never mixes GPU and CPU output.
- A GPU is not always faster. An integrated graphics chip can be slower than the CPU.
- Restart the app after installing a new driver or adding a graphics card. If a saved card cannot be found, it shows as Saved GPU · not found and films run on the CPU.
Crop localizer
With Crop localizer On (the default), SpineContour searches the film for the lumbar spine and crops to it before measuring. Full-spine films need this, because otherwise the lumbar vertebrae are too small for the models.
Turn it Off for lumbar-only films that are already framed around the spine. The models then run once on the whole image, after trimming broad black borders, which saves the repeated search. The S1 endplate and femoral heads are still detected.
Off is for lumbar films only
With Crop localizer Off, a standing full-spine film is measured without being cropped, and the lumbar anatomy can be too small to read reliably. Switch it back On before processing full-spine films.
Crop method
How the region to measure is found on full-spine films, and on lumbar films while Crop localizer is On.
- Crop search (the default) checks many candidate regions across the film.
- Trained model runs a region detector once to propose the cervical and lumbar crops. A proposal still has to pass the same landmark checks. If it is missing or fails them, that region falls back to Crop search.
Each result records which method supplied the crop. Cervical-only films do not use a crop method. The region detector always runs on the CPU, whatever the Processor setting.
Toolbar removal
Off by default. When On, SpineContour looks for a PACS toolbar strip along the bottom of a screenshot and trims it off before measuring, so the buttons are not mistaken for anatomy.
- It is cautious. Only bottom rows are removed, and only when the strip clearly looks like a toolbar. Anything uncertain is left as it is.
- It only applies to 8-bit grayscale screenshots. DICOM files and higher bit-depth images are left alone.
- Your file is never changed, and calibration still reads the original image.