CHD segmentation

Two pairs of paediatric studies, each a heart with a congenital defect beside a heart without one, and each outlined by the people who published it. Scrolling either study moves its partner to the same level, so the two can be read against each other, and drawing a box round a structure on one names it on both. The CT pair is a tetralogy of Fallot. The MR pair comes from a dataset that records defects one at a time without naming syndromes, so it is a septal defect rather than a tetralogy.

Reading the study…

On CT

ImageCHD, two children scanned on the same scanner at the same 0.24 mm voxels, seven structures outlined.

Tetralogy of Fallot (ImageCHD case 1046)

Healthy control (ImageCHD case 1080)

Press Identify, then drag a box round a structure on either study. The other names the same structure on its own slice.

On MR

HVSMR-2.0, two children of thirteen and fourteen, eight structures outlined. This dataset records defects one by one and never names a syndrome, so neither does this pair: it is a septal defect beside a heart with none, not a tetralogy.

Septal defects (HVSMR-2.0 case 14, age 13)

No recorded defect (HVSMR-2.0 case 8, age 14)

Press Identify, then drag a box round a structure on either study. The other names the same structure on its own slice.

Questions

Four published studies in two pairs, read side by side, to see whether the browser can hold a comparison. Not a reading of these images and not for diagnostic use.

What is a tetralogy of Fallot?

Four things found together, and they follow from one another rather than arriving separately. The wall between the ventricles is incomplete, so the two sides of the heart are not kept apart. The aorta sits over that gap instead of over the left ventricle alone, so it receives blood from both. The way out of the right ventricle towards the lungs is narrowed. And the right ventricle, pushing against that narrowing, thickens.

The outlines on this study name normal anatomy: a ventricle is a ventricle whatever it is doing. What makes the case a tetralogy is how those structures sit relative to one another, which is something to look for by scrolling rather than something any single outline records.

How do I drive the viewer?

The wheel moves through the stack, and the arrow keys do the same once the image has focus; the heart occupies the middle of the series. Dragging with the left button changes the window, and the presets are true Hounsfield values, so Soft tissue and Lung mean here what they mean on a workstation. The right button zooms, the middle button pans, and Reset puts both back.

Identify is the one control that is not a view setting. Press it, drag a box round a structure, and the structure is named beside the image with its outline drawn over the anatomy.

Why does scrolling one study move the other?

Because the point of the page is the comparison. The two children are different sizes and their scans start and stop at different levels, so the link maps between the outlined hearts rather than between slice numbers: it holds the level through the heart and drifts outside it. Window, zoom and pan stay separate, so each study can be set up on its own.

What is the control, and in what sense is it healthy?

ImageCHD case 1080, one of the twenty cases the dataset lists no congenital defect against. That is what the table records, not a reading of the images, and the child was still scanned for some clinical reason. So it is a comparison rather than a certificate of a normal heart.

The pair was chosen to match: the same scanner, the same 0.24 mm voxels and the same 0.75 mm slices, so a difference between them is a difference in the hearts rather than in how they were scanned.

Why is the MR pair not a tetralogy?

Because its dataset does not say that any case is one. HVSMR-2.0 records defects one at a time, as a ventricular septal defect, a double outlet right ventricle, a transposition, an absent pulmonary artery, and it never groups them into a named syndrome. A tetralogy is exactly such a grouping, so calling one of these children a tetralogy would be a diagnosis made here rather than one published there, which is not something a page like this should do.

What the MR pair does show is the septal defect a tetralogy shares, in a child of thirteen, beside a heart of fourteen with nothing recorded against it. The two were picked to be close in age for the same reason the CT pair was picked off one scanner: so that a difference between them is more likely to be in the hearts.

Why do the two pairs have different controls and different structures?

Because the two modalities do not measure the same thing. A CT number is in Hounsfield units, which mean the same on any scanner, so Soft tissue and Lung are real presets with fixed values. MR signal has no such scale: the same tissue can be any brightness on a different sequence, so the MR pair offers the window the file was written with, the full range of the image, and a narrow one, and nothing claims to be a tissue preset.

The outlines differ too. ImageCHD labels seven structures including myocardium; HVSMR-2.0 labels eight, leaving out myocardium and adding the superior and inferior venae cavae. Each pair names what its own annotators drew.

Where do the images and the outlines come from?

The CT pair is ImageCHD, 110 congenital heart CT studies published by Xiaowei Xu and colleagues and described in their MICCAI 2020 paper. Released under Attribution-NonCommercial 4.0 International (CC BY-NC 4.0); this page is non-commercial and carries the attribution the licence asks for.

The MR pair is HVSMR-2.0, 60 cardiovascular MR scans of children with congenital heart disease from Boston Children's Hospital, published by Danielle Pace and colleagues in Scientific Data in 2024 under Attribution 4.0 International (CC BY 4.0).

Every outline on this page is the published one, re-encoded rather than redrawn.

How did NIfTI volumes become something the viewer can read?

Both datasets ship NIfTI, and each has its own converter because they do not present the same problems. scripts/nifti-to-dicom.py handles ImageCHD, whose affine is the identity: it restores the voxel spacing from the dataset's own spreadsheet, settles the orientation from the anatomy, and shifts the stored values into Hounsfield units so the CT presets mean what they say. scripts/hvsmr-to-dicom.py handles HVSMR-2.0, which carries a real affine, so the geometry is read from the file and nothing is rescaled, because there is no scale to rescale onto. No patient dates are carried over by either.

The page then runs the same Cornerstone3D engine as the other viewers, with its own copy of the bundle and the codec binaries, because the Emscripten decoders resolve their .wasm against the page that loads them.