Thank you!
I'm fully aware of the problems regarding the CSF between gray matter and the skull. Especially in these low-dose CTs the partial volume effect is very intense in this area.
In your opinion, would it help the segmentation if I added a second channel, where I gave in input a segmented image of only the bone? (as my percentage of patients where the segmentation fails generally have the c4 class completely wrong)
Best regards,
Luca
________________________________________
Da: John Ashburner <[log in to unmask]>
Inviato: mercoledì 16 novembre 2016 15.06.21
A: PRESOTTO LUCA
Cc: [log in to unmask]
Oggetto: Re: [SPM] Tweaking segment for CT images
This is something I've been thinking about, but it hasn't been implemented in SPM yet. Preliminary work for learning tissue intensity priors was done in:
Blaiotta C, Cardoso MJ, Ashburner J. Variational inference for medical image segmentation. Computer Vision and Image Understanding. 2016 Apr 11.
As most people use T1-weighted scans, having intensity priors for these should bring some additional robustness. Much more informative priors should be achieved for CT, as Hounsfield units should not vary from scanner to scanner (unlike the contrast in T1-w images). Problems do arise with thick sliced CT though, because partial volume effects often mean that CSF outside the brain is not properly visible (through mixing with the much higher intensities from the skull). Major problems can be encountered because this CSF and that in the ventricles is supposed to have the same intensity distribution, when in reality it doesn't.
Best regards,
-John
On 16 November 2016 at 13:11, PRESOTTO LUCA <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear experts,
I recently tried using the "segment" function in spm12 on extremely low dose CT images and I was impressed by how well it works, despite the terrible image quality. Obviously the grey matter segmentation isn't resolute at all, gyry aren't identified. You only get a single large thing including all of the area where grey matter is present. But it does find the borders of this area correctly, as compared to MRI. And this results in a very good normalization!
I've done some preliminar tests on 10 cases and the algorithm worked more than perfectly in 8 cases and failed markedly in 2 (segmenting as bone also part of the grey matter and of the outside tissues), resulting in a wrong normalization.
Therefore, as in a CT images one could provide very good priors for the intensities of bone, air, CSF and outside tissue even in extremely noisy situations I was wondering... Is there an easy way to do this? I'd think it could easily end up with a 100% success rate!
Thank you,
Luca
Rispetta l’ambiente: non stampare questa mail se non è necessario.
Respect the environment: if it's not necessary, don't print this mail.
IL TUO 5XMILLE AL SAN RAFFAELE DI MILANO
Devolvi il tuo 5 per mille all’Ospedale San Raffaele: perché al centro della Ricerca ci sei TU. CODICE FISCALE: 07636600962, nel riquadro RICERCA SANITARIA. Non c’è cura, senza ricerca. Non c’è ricerca, senza il tuo 5xmille. Scopri come su http://www.5xmille.org
[http://www.5xmille.org/img/731x129.jpg]
Rispetta l’ambiente: non stampare questa mail se non è necessario.
Respect the environment: if it's not necessary, don't print this mail.
IL TUO 5XMILLE AL SAN RAFFAELE DI MILANO
Devolvi il tuo 5 per mille all’Ospedale San Raffaele: perché al centro della Ricerca ci sei TU. CODICE FISCALE: 07636600962, nel riquadro RICERCA SANITARIA. Non c’è cura, senza ricerca. Non c’è ricerca, senza il tuo 5xmille. Scopri come su http://www.5xmille.org
[http://www.5xmille.org/img/731x129.jpg]
|