Résumé :
|
Introduction: Accurate muscle size quantification is essential to investigate muscle function related to exercise adaptation, ageing and neuromuscular pathologies. Currently, MRI muscle volume quantification is the gold standard; however this technique requires a time-consuming manual delineation of the muscles of interest. Little work has been done so far on automated segmentation of specific muscles. In this work we propose a fully-automated method to segment region of interest on limb muscle images. We evaluated this method for the segmentation of rat gastrocnemius and plantaris muscles. Methods: The automatic segmentation method is based on multi-atlas method. Each image is segmented using a set of atlases; the segmentations produced are then combined by a voting procedure. To further enhance this process, we generate artificial atlases by shifting the atlases along Z before the non-rigid registration and adding their segmentation in the vote. The leg images of 12 rats (2 months old wistar Han) were obtained using MRI from the ankle to the knee with a Rare sequence (TR=2000ms TE=16ms, FOV=30x32mm, 18 slices). Gastrocnemius and plantaris muscles were manually segmented on 5 other rat images which were used as atlases. The method has been tested using 1, 3 and 5 atlases, with and without the Z-shifting step. The muscle volume error (MVE) was the relative differences between the volume of automatic and manual segmentation. The relative overlap (RO) was the volume of the intersection between manual and automatic segmentation to their union volume ratio. Results: The MVE was reduced using 3 atlases as compared to 1, but no difference was observed when 3 or 5 atlases were used. Considering this parameter, the benefit of the Z-shift procedure was surprisingly not clearly visible (Table 1). However, regarding the RO we found an improvement using Z-shift and in respect with the number of atlases. More specifically the RO obtained with Z shift and 1 atlas was similar to this obtained with 3 atlases and no Z-shift (similarly, Z-shift, 3 atlas no Z-shift, 5 atlases). Conclusion: Our results sho wed that the present method is reliable to quantify the whole muscle volume of different limb muscles with a minimal operator cost. The Z-shift procedure introduced in this work enables a further improvement and a reduction of the atlases number in the manual database. As this method is fully automated no intra- and inter-operator variability is introduced during the processing.
|