Manual Versus Automated Segmentation

As mentioned at the beginning of this chapter, detection, localization, diagnosis. staging, and monitoring treatment responses are crucial procedures in clinical medicine and oncology. Early detection and localization of the diseases and accurate disease staging could lead to changes in patient management that will impact on health outcomes. Noninvasive functional imaging is playing a key role in these issues. Accurate quantification of regional physiology depends on accurate delineation (or segmentation) of the structure or region of interest (ROI) in the images. The fundamental roles of ROI are to (1) permit quantitation, (2) reduce the dataset by focusing the quantitative analysis on the extracted regions that are of interest, and (3) establish structural correspondences for the physiological data sampled within the regions.

The most straightforward segmentation approach is to outline the ROIs manually. If certain areas in the images are of interest, the underlying tissue time-activity curve (TAC) can be extracted by putting ROIs manually around those areas. Approaches based on published anatomic atlases are also used to define ROIs. The average counts sampled over voxels in the region at different sampling intervals are then computed to compose the TAC for that region. The extracted tissue TACs are then used for subsequent kinetic analysis (Chapter 2 of Handbook of Biomedical Image: Segmentation, Volume I).

In practice, selection of ROI is tedious and time-consuming because the operator has to go through the dataset slice by slice (or even frame by frame) and choose the most representative ones from which 10-40 regions are carefully delineated for each imaging study. Needless to say, manual ROI delineation is also operator dependent and the selected regions are subject to large intra- and interrater variability [8, 9]. Because of scatter and partial volume effects (PVEs) [10], the position, size, and shape of the ROI need careful consideration. Quantitative measurement inaccuracies exhibited by small positional differences are expected to be more pronounced for ROI delineation in the brain, which is a very heterogeneous organ and contains many small structures of irregular shape that lie adjacent to other small structures of markedly differing physiology [11]. Small positional differences can also confound the model fitting results [12,13]. To minimize errors due to PVEs, the size of the ROI should be chosen as small as possible, but the trade-off is the increase in noise levels within the ROI, which maybe more susceptible to anatomical imprecision. On the other hand, a larger region offers a better signal-to-noise ratio but changes that occurred only within a small portion of the region maybe obscured, and the extracted TAC does not represent the temporal behavior of the ROI but a mixture of activities with adjacent overlapping tissue structures. Likewise, an irregular ROI that conforms to the shape of the structure/region where abnormality has occurred will be able to detect this change with much higher sensitivity than any other geometrically regular ROI that may not conform well. In addition, manual ROI delineation requires software tools with sophisticated graphical user interfaces to facilitate drawing ROIs and image display. Methodologies that can permit semiautomated or ideally, fully automated ROI segmentation will present obvious advantages over the manual ROI delineation.

Semiautomated or fully automated segmentation in anatomical imaging such as CT and MR is very successful, especially in the brain, as there are many well-developed schemes proposed in the literature (see surveys in [14]). This may be because these imaging modalities provide very high resolution images in which tiny structures are visible even in the presence of noise, and that four general tissue classes, gray matter, white matter, cerebrospinal fluid (CSF), and extracranial tissues such as fat, skin, and muscles, can be easily classified with different contrast measures. For instance, the T1- and T2-weighted MR images provide good contrast between gray matter and CSF, while T1 and proton density (PD) weighted MR images provide good contrast between gray matter and white matter. In contrast to CT and MRI, PET and SPECT images lack the ability to yield accurate anatomical information. The segmentation task is further complicated by poor spatial resolution and counting statistics, and patient motion during scanning. Therefore, segmentation in PET and SPECT does not attract much interest over the last two decades, even though there has been remarkable progress in image segmentation during the same period of time. It still remains a normal practice to define ROIs manually.

Although the rationale for applying automatic segmentation to dynamic PET and SPECT images is questionable due to the above difficulties, the application of automatic segmentation as an alternative to manual ROI delineation has attracted interest recently with the improved spatial resolution of PET and SPECT systems. Automatic segmentation has advantages in that the subjectivity can be reduced and that there is saving in time for manual ROI delineation. Therefore, it may provide more consistent and reproducible results as less human intervention is involved, while the overall time for data analysis can be shortened and thereby the efficiency can be improved, which is particularly important in busy clinical settings.

0 0

Post a comment