I asked the hospital for my MRI on a disc and got 1,189 image files. The scan was for numbness down one side, and covered my brain, neck and lower back in one sitting, including a contrast injection. The files were in DICOM (Digital Imaging and Communications in Medicine), the medical-imaging standard that stores images alongside scan settings and patient information.
The viewer here contains 515 selected slices in 16 series. These are grayscale JPEG exports with a fixed contrast window, the range of signal values mapped to black through white. They don’t preserve the full signal range of the DICOM files. Pick a region and series, then scroll, drag or scrub the slider. Play runs through the stack.
Why the same tissue looks different
MRI detects radio signals from hydrogen nuclei, mainly in water and fat. Magnetic-field gradients encode where those signals came from, and a computer reconstructs the image. It doesn’t measure one little box of tissue at a time. Brightness depends on the pulse sequence, its timing and the tissue’s magnetic properties, rather than simply its density. NIH’s MRI introduction and Siemens’ sequence guide explain the signal and encoding.
A sequence is a programmed pattern of radio pulses and gradients. T1 and T2 refer to two ways the signal relaxes after excitation. In ordinary T1-weighted images, fat is usually bright and spinal fluid dark. In T2-weighted images, fluid is usually bright; fat can still be bright unless suppressed. T2 is not a reversal of every shade in T1. Lauder and Driscoll’s sequence table compares them.
FLAIR suppresses spinal fluid; STIR and the labels marked FS suppress fat. SWI emphasizes differences in how tissues affect the local magnetic field, useful for seeing veins and iron-rich tissue. DWI is sensitive to the motion of water molecules, and MEDIC combines several signal echoes. Post-contrast means images taken after the injection. Axial slices cut across the body; sagittal slices divide left from right; coronal slices divide front from back. These are different acquisitions, so slice 12 in one series needn’t match slice 12 in another. Sequence guide; original SWI paper.
What the models can show
The brain below is a processed surface made from T1 images. Its build used neural-network estimates between slices, removed the surrounding skull and scalp, filled the ventricles and closed deep grooves before smoothing the surface. Those edits make it easier to display as a solid object, but alter its shape. It isn’t an untouched cortical surface.
The second model is a rough torso envelope from a localizer, a scan used to plan the examination. It doesn’t trace individual vertebrae. MRI can show marrow and abnormalities in bone, as well as discs, nerves and the cord; the limits of this model aren’t the limits of spine MRI.
Drag to rotate and scroll to zoom. This viewer draws two finished triangle meshes, not a volume you can cut through to expose the original signal.
How the web version was made
- Keep compatible slices together.
DICOM records image position, orientation and pixel spacing. These locate a slice in the body. Slice thickness and the distance between slice centers are separate measurements. A volume needs a consistent acquisition and geometry; stacking different planes or contrasts together would mix different measurements of the same tissue.
- Export the slice sheets.
Each selected series is packed into a JPEG sprite sheet, a grid of slice images. The JavaScript crops one tile onto the canvas. Most tiles are 448 by 448 pixels; the SWI tiles are 240 by 256. These are display dimensions, not measurements of anatomical resolution. The exports contain no physical slice-spacing fields.
- Estimate between slices.
For the brain build, I trained a small network on degraded versions of the scan’s sharper slices, then used it to estimate detail in the thicker direction on a roughly 0.9 mm grid. That is output spacing, not proof of 0.9 mm anatomical resolution. Self-super-resolution research makes the distinction: resampling adds grid points, while missing detail has to be estimated. Correct coordinates allow a 3D display even when the original voxels aren’t cubes.
- Make the surface.
HD-BET estimated a brain mask, separating brain tissue from its surroundings. I filled the ventricles and closed deep grooves, then used marching cubes to turn the mask boundary into triangles. Smoothing and simplification left 120,000 triangles in the brain mesh. The NiiVue WebGL2 library draws it in the browser. The original images and training logs aren’t included in these web assets, so the finished mesh alone can’t establish the accuracy of the reconstruction.
What 1.5 tesla tells us
My scan used a Siemens MAGNETOM Sola, a 1.5-tesla system. The number is the magnet’s field strength. Field strength affects the available signal, but scan time, motion, coils and sequence settings also matter. Clinical MRI uses both 2D and 3D acquisitions. A JPEG export or a smooth model isn’t enough to rank this examination against a 3-tesla scan, or to explain why a particular protocol was ordered.
I haven’t tried to diagnose anything from these images. The JPEGs have fixed contrast and compression, and the models include estimated and edited detail. A clinical reading belongs to the original study and the radiologist’s report.