Comments on manual segmentation methods
Manual segmentation methods use the pattern recognition abilities of the human visual system. The human brain remains unbeatable for pattern analysis tasks. These days, it is common to use a computer to record the manual delineations, a task at which computers remain unbeatable. Manual image analysis is perfect for small-scale image analysis since it requires little by way of "programming". However, it becomes tedious, costly, and impractical for large-scale studies. For such studies, automated segmentation is preferable, since the programming effort gets amortized over a large number of images.
The main limitation with manual analysis is subjectivity. Different people may come up with different segmentation results - this is known as inter-subject variability. Interestingly, the same person can produce different segmentations of the same image at two different times - this is known as intra-subject variability. Automated systems offer a measure of objectivity, and repeatability.
As noted above, the human visual system remains unbeatable for pattern analysis. On the down-side, the human hand is remarkably unsteady. Some computer-assisted systems attempt to compensate for this hand unsteadiness (and slowness) by using real-time image analysis algorithms (e.g., digital scissors and contour fitting algorithms).
The human visual system has limited 3-D capabilities - it is restricted to stereoscopic viewing of a scene. It cannot visualize volumetric data.
Many attempts have been made to make human image analysis faster. One major approach is known as unbiased stereology. These methods work by providing systematic ways to subsample the image data, and extrapolate from manual analysis of subsampled regions. As the name implies, these methods strive to minimize the statistical bias associated with the subsampling. In order to achieve this goal, several assumptions are made. One such assumption is that the tissue is homogeneous - this is rarely the case. Even when the achieved bias is low, it is possible for the variance to be high. Lowering the variance requires analysis of many more tissue samples. Indeed, some stereologists advocate extraction of small numbers of measurements across large numbers of specimens. Finally, stereological methods are poorly suited to multi-dimensional data and analysis of associations.
This is not to imply that stereological methods are to be abandoned. There are cases when automated analysis can be carried out on a stereologcally sampled set of images for very large-scale problems. When such experiments are designed one must pay attention to bias and variance.
