01. Automated organ segmentation correction
Automated tools can now segment organs and muscles quickly, and the 3D result often looks convincing. But errors hide at the boundaries, where small over- or under-segmentations skews volumes, lesion quantification and radiomic feature extraction. In the example below, a clean-looking 3D model turns out to miss large parts of several structures when checked against the native images.
I review and correct automated output from proprietary or open-source tools to a standard fit for quantitative research and clinical use, with a treating clinician's review as a final check where appropriate.
3D visualisation vs. native images: undersegmentation of the left iliopsoas, gluteus maximus, deep back muscles, bladder and small bowel by automated algorithm.
02. Brain segmentation correction
Neuroimaging pipelines can pass QC while still including structures that don't belong, such as dura mater, falx cerebri or tentorium cerebelli, or overestimating gray matter in frontal and temporal lobes. Standard metrics rarely catch this.
I manually refine these boundaries against the native scan. Often the pipeline is already largely correct. The work is fixing the last few percent that matters. How much correction is needed depends on the application: large-cohort studies can tolerate errors that individual-level or clinical analysis can't. For large cohorts, I also offer QC without correction: reviewing each segmentation, flagging problems and identifying which scans need attention or should be excluded.