In my research I have analysed structural volumes in MRI scans by manual tracing measures and correlated against automated measures, with particular interest in subcortical grey matter structures. As I am still a student it is interesting to gauge the overall need for the skill in the international community, either as a diagnostic tool or as an inter-rater reliability measure.
Excellent opinions ...I think all in respect to the general goal behind the question...
I think there's always going to be a need for expertise in that area. As new automated anatomical parcelation techniques are developed, it would be prudent to have third-party experts provide a critical assessment of the efficacy of these methods. Moreover, sometimes, automated methods just fail. The only alternative to a manual approach would be to give up on those datasets
I would certainly agree with Chris. Traditionally, manual methods have been treated by some as a 'gold standard', and although there are unquestionably some advantages to human decision-making (such as taking account of anatomical variants - e.g. paracingulate - or increased variance in atrophy which may be taxing for automated methods), it could also be argued that rater drift and similar issues of non-systematic error make automated methods more attractive. When one also has to consider the degree to which MRI landmarks might actually correspond to your ROI's cytoarchitectural and hodological characteristics (particularly for cortical areas), you could argue that even the most 'accurate' manual parcellation might not be an entirely true reflection of the underlying biology, making any minor variations between methods academic. (However, in my limited experience, inter-method variations do not always tend to be minor...). I'd be interested to hear how your comparisons have turned out.
In large-sample studies with increased power to detect small effects reliably, the simple issue of resource has an unfortunately large role to play. Manually tracing hundreds of MRIs across time points is a massive task in terms of person-hours. The availability of automated software to do this at relatively lower cost (combined with the fact that results using such methods are eminently publishable) makes a persuasive case for the automated route. However, the skills you acquire when applying manual tracing are highly useful in an automated setting too. Understanding the methods and rationale upon which automated atlases are predicated is useful in guiding your atlas selection (or development of your own), interpretation of the outputs and subsequent analyses. Being well-versed in manual segmentation also has benefits when running visual checks on the outputs (something that really should be done to at least some extent).
I would also say that although some automated methods do a spectacularly complex task pretty well, I might still be more convinced by a report that used (a reproducible and well-justified) manual tracing protocol if in a small sample of participants for whom automated methods have not been adequately tested or developed. And let's face it, these are often some of the more interesting groups to research!
Thank you Chris and Simon,
That definitely sets my mind at ease. I'm currently looking at the effects of HIV-1 infection on brain development. It includes a sample of perinatally HIV-1 infected children and uninfected controls (n=78) with 11 sub-cortical structures of interest. Like you said it was extremely time consuming, in total it was approximately 800hrs of tracing. I have finished manually tracing all the scans, statistics have been computed including correlating the manually acquired structural volumes with automated measures. I'm busy writing up my thesis and will be writing up for publication shortly after if you're interested in the results.
Besides seeing the effects of HIV-1 infection regarding manually acquired volumes, and correlating manual vs automated, I too wanted to see if automated measures were as sensitive as the manual in picking up similar effects caused by HIV-1 infection using the automated measure derived volumes. preliminary results are quite interesting.
Much is made, in the clinical situation, of relative tonsil and adenoid size in children. "Obstruction" of the airway is inferred from clinical examination... which has aspects of the uncertainty principle. A safe, practical, inexpensive way to document the volumetric relationships of tonsils and adenoids would be a significant contribution to airway analysis in children.
I agree that sometimes the automated methods fail leading to miscalculations and to unreliable results. Nonetheless I would humbly say that manual approaches are quite user-dependent and need caution. And of course are very time-consuming. I think it depends on the margin of error "allowed" considering the expected effect size. Maybe for research new automatic segmentations are enough but for clinic a manual tracing is preferable. Semi-automatic methods are probably the best choice in between.
For neurosurgical use, the manual segmentation is too time consuming prior to emergency operation. Practically, estimation of the size of lesions seeing the three horizontal, coronal and sagittal view is sufficient. For animal experiment, planimetry of HE stained sections of constant thickness enlarged on the papers by CP scanner is more practical.
- Badges
- Science topic
- Similar topics
- Biological Science
- Neuroscience
- Neuroanatomy