There are a few options.

• Downsample your image. If the region you're trying to segment is fairly large and homogenous, this won't impact your results too much.
• You can try a region growing algorithm, which will work well if the edges of your shapes are reasonably well delineated and your shapes are somewhat homogenous. As an example, can use Mevislab's RegionGrowingMacro for this, but this is probably also available in software like ImageJ. 
• If your threshold or region growing gives somewhat desirable results but can't differentiate between large or small shapes, try a connected component analysis. This will sort your shapes by volume. If your large and small shapes are still connected slightly, run an erosion followed by a dilation morphological operator before connected component analysis. This can all be done in Mevislab.
• If you want a more fancy method, consider ITK-Snap's Snake method (see link). What it does is it takes a starting sphere as an input and expands using a model based on natural physics it until it finds 'edge' voxels. Also read up a bit on active contour models. 
• ITK-snap also has a 'similarity brush' tool you might be interested in, but it'll require you to go through every slice manually.

http://www.itksnap.org/docs/viewtutorial.php?chapter=TutorialSectionIntroductionToAutomatic

http://www.mevislab.de/

Thank you all for the response. I will try implementing the methods and see how it goes.