Case study: visual QC as complementary evidence

Every concept in this chapter is invisible in an array and visible in a viewer. A mask that is one slice off the tumour, a resampling that shifted the boundary, a registration that quietly misaligned — these are things you see, then measure. This closing case presents a real visual-QC pass on RIDER data as a scripted, auditable procedure. The procedure makes clear what looking can and cannot certify.

A mask can look right and be geometrically wrong

Open a CT with its tumour mask. The mask is a coloured overlay roughly where the tumour is. Is it exactly on the tumour, in every plane, at every slice? You cannot tell from one axial view. You need to walk the volume in three planes, toggle the mask, and compare source against processed — and then record a verdict.

The complementarity cuts both ways. Visual inspection catches the gross failures that automated checks can miss when their assumptions are wrong: an overlay that sits one slice off, a boundary that moved after resampling, a registration that misaligned anatomy while every tag looked plausible. But eyes cannot certify what the geometric checks certify: unaided appearance cannot certify a small metadata difference — whether a change such as 0.1 mm of spacing is even perceptible depends on the accumulated displacement, the field of view, the task, the display, and the tools, and metadata panels, rulers, or automated checks can expose such differences directly — and no overlay proves that source references or hashes are intact. Visual QC complements geometric and provenance checks; it does not replace them.

The setting: 3D Slicer as a geometry viewer

3D Slicer is the tool used here for seeing geometry because it reads the affine for you and renders true multi-planar views. The 3D Slicer companion maps the verified Witowski case material to these concepts; the points that matter for this case are three. Scrolling coronally through an axially-acquired volume is resampling in disguise — Slicer reslices the volume on the displayed plane, interpolating from the full image geometry: the complete affine, with its origin, spacing, and direction, not the direction matrix alone (the resampling section). DICOM import is organised around the Patient/Study/Series hierarchy and loaded by plugin-specific interpretation: SeriesInstanceUID identifies a series, and the chosen plugin decides how the files become volumes — it may offer several interpretations or subseries, and may regularise inconsistent acquisition geometry — after which Slicer works from the spatial metadata and referenced objects available; grouping alone neither resolves the frame of reference nor automatically validates consistency (Chapter 3, the shared-space section). And window/level maps HU→display, recalling the slope/intercept of Chapter 3.

The case: a scripted QC pass on RIDER data

The pipeline under study exports QC bundles — source CT, source mask, processed CT, processed mask — together with a Slicer launcher that opens them in a 2×3 layout (rows: source vs processed; columns: axial/coronal/sagittal) with the bundle’s lung window (W≈1500, L≈−600 in HU display). The committed bundle tooling also records provenance, geometry, and hashes, so what a reviewer sees is tied to what the pipeline actually produced. The procedure proceeds in five steps.

Load the study. Drag the RIDER CT series into Slicer; confirm the lung window and that axial/coronal/sagittal reformats are consistent — they are all computed from one geometry, so inconsistency here is itself a finding.

Overlay the tumour mask. Load the segmentation (from the SEG or the exported .nrrd). Toggle visibility on/off. Confirm the mask sits on the tumour, not one slice off. If it is off, that is an observation, not a diagnosis: the offset could come from wrong source selection or linkage, a registration step, transform direction, segmentation provenance, frame mapping, or the complete grid geometry. Investigate that chain first, then re-run the shared-space checks.

Inspect the 3D model. In Models / Segmentations, toggle the 3D view. A boundary that “shimmers” or a slab-like silhouette are non-specific appearances, not diagnoses: candidate causes to investigate include view interpolation, contour stair-stepping, anisotropic sampling, the segmentation’s own shape, thresholding, rendering, motion, or genuine anatomy (the resampling section, Chapter 1’s reconstruction case). Appearance alone does not license a diagnosis of partial volume, label noise, or a slice-thickness artefact.

Compare source vs processed. Open the processed image and processed mask (post-resample). Walk through slices and ask: did the tumour shape survive resampling? Is the boundary in the same physical place? Toggle between source and processed to catch a mis-registration.

Record a verdict. Use the pipeline’s review recorder: PASS / FAIL / NEEDS_INVESTIGATION plus a reason. This is what makes the pass auditable: an unrecorded inspection cannot count as completed evidence.

Toggling the mask off removes the coloured overlay and re-reveals the underlying image; it does not make the tumour disappear from the CT. A plausible overlay across all three planes, at every slice, is bounded evidence — evidence against gross visible defects such as an obvious shift, rotation, or clipping — not a certificate that the affine, the slice ordering, and the CT↔SEG grid agreement are all correct at once. And when something looks wrong, the plane or slice where it fails does not uniquely identify which upstream check is at fault — affine, ordering, shared-space geometry, and resampling each remain candidates to re-run.

What the verdict means

A recorded PASS in this case means that a human walked the volume in three planes, compared source against processed, and found no gross geometric failure — with the bundle’s hashes and geometry record available beside the images. It does not mean the mask is provably on the CT grid, that the resampling was optimal, or that the provenance is complete; those are the work of the geometric checks, the reference validation, and the manifest, respectively. The honest reading of visual QC is the one in this chapter’s opening: it is complementary evidence — strong at catching what assumptions missed, silent about what eyes cannot resolve.

The chapter set out from a mask that looked right and sat in the wrong space. The intervening sections built the tools that make such a failure detectable rather than fatal: the affine that locates a voxel, the slice normal that orders a stack, the grid comparison that decides shared space, the representations that carry or destroy provenance, and the resampling rules that keep interpolation from inventing measurements. Each tool turned one silent failure mode into an explicit check, and this case exercised them together on real data.

Geometric validity is where this part of the book ends, and it is not yet a measurement. A volume can be perfectly placed, correctly ordered, and faithfully resampled, and the number a feature extracts from it can still be a poor answer to a poorly posed question. The next chapter opens that problem: a number is the output of a measurement system, and the image is only one part of it.

Sources

• 3D Slicer User Guide — DICOM and Volumes modules: DICOM hierarchy, plugins, import, and loading; volumes, grid geometry, slice views, interpolation, window/level, and overlays. (slicer.readthedocs.io)
• RIDER lung CT collection (Version 3) — Zhao B, Schwartz LH, Kris MG, Riely GJ, The Cancer Imaging Archive, doi:10.7937/K9/TCIA.2015.U1X8A5NR. (doi.org)
• TRACE-CT visual-QC evidence at commit 7aaf308c76ed2718fd0b364d1cde856969711b72 — committed RIDER visual-QC aggregate, bundle schema, Slicer bundle launcher, and review recorder. (trace-ct commit 7aaf308c)