MARS - Models of Appearance, Registration and Segmentation - Registration Without Ground-Truth Solution

T has been rather long since an update was last posted. I am pleased to have stitched together some results from one month of computer labour. The joint and averaged results prove that our endeavour has been a smashing success.

Below lie two graphs which illustrate our ability to evaluate registration without any ground truth, e.g. data labels which had been manually placed a priori. In other words, we can evaluate registration from just the raw images and the registration's deformation fields. The method relies on models of appearance and the shuffle distance, as previously and quite repeatedly described. The method was also shown to be correlated with results that are in fact based on the ground truth, namely the manually annotated brains from MGH. More information on these results is available in a previous write-up.

Our measures increase monotonically as mis-registration in brain data increases (click to enlarge)

As misregistration increases, so do the Specificity and Generalisation ability of an automatically constructed appearance model.

The implementation we took advantage of had been written in C++, but it was derived from my earlier and equivalent implementation in MATLAB, soon to be published as an Open Source project.

Meanwhile, I am beginning to write up my thesis, which will cover all my work rather concisely and yet at a greater level of depth.

I recently noticed that a few dozens of people syndicate this site and/or have subscribed as members. Thank you all very much for showing interest.