Evaluation of head segmentation quality
-
Do you have a ground Truth segmentation to compare it with? We compared it to manual segmentation for three different datasets (four IXI subjects, the MIDA model, and a third dataset with six subjects from a collaboration) and found it to be favorable wrt accuracy (Dice, Hausdorff, mean distance, etc) to alternative methods.
-
No, unfortunately I don't have a ground truth. The problem is that I would like to run simulations both with models obtained using the S4L head segmentation toolbox (the one already available in v7, not the new one in v8) and models in Virtual Family. My concern is that results obtained from these two kinds of models are not comparable if I can't check for accuracy and ensured to be of comparable quality. I would like to ensure consistency in segmentation.
-
Most of the design choices in the head40 segmentation are similar to ViP V3 models. Some differences include
- ViP: the CSF is not modeled as closed layer, here you could disable "Close CSF" in head40. ViP models have a closed dura, so here you should select "Add Dura".
- ViP has tendon ligaments along posterior ridge of vertetrbrae: this is not included in head40
- ViP have no galea on the skull surface (tendon), MIDA and head40 does
- ViP separates teeth
- head40 has "Other_tissue". This is mostly fat and would be either SAT or Fat in ViP (which have same properties)
- ViP: arteries and veins are more detailed & accurate
- other differences in segmentations should not result in systematic differences in material properties
-
head16 includes much fewer tissues, and these are less accurate than in head30 or head40. head30 has similar accuracy (compared to head40) but lumps muscle, tendons, tongue, glands, and other labels into "Other_tissue". head40 is most comparable to ViP segmentations (with differences listed above)
