DICOM Lens — entertrainment

PROJECT REVIEW · 18 October 2026

Inside the project

Prepared with the project’s development agent from source and retained evidence. This is a project-owner assessment, not an independent customer review. Dated checks were not rerun for this website.

The Lens is the 3D stage of the NeuroAtlas DICOM workstation. A study's segmented structures sit as persistent layers you can show, shade, isolate, cut through and fly around. Since October it can do more than show anatomy: you can pick a structure and pull it out to examine it enlarged and turned. You can call it 'this' and ask the built-in assistant about it; the assistant answers only what the scan supports, such as a Hounsfield-unit profile on calibrated CT, and refuses the rest with a reason. You can also isolate or cut it by voice or gesture. Vessel bifurcations found by the vessel analysis appear as markers you can see through the vessel wall. A displaced or enlarged copy is display-only: every measurement stays in the scan's own millimetres. It is a research and engineering tool, explicitly non-diagnostic, and it runs offline on macOS.

The 18 October 2026 website demonstration uses actual outputs from the AssistantApp offline DICOM decoder on a newly generated synthetic CT phantom. The 176 × 176 × 72 volume contains six known HU values. Decoded voxel counts and means for all six classes, and the whole-volume mean, match the independently generated ground truth. The browser viewer selects retained axial, narrow-window, coronal MPR and 60 mm maximum-intensity-projection images. It does not run the decoder in the browser and does not capture the native Lens interface. This bounded engineering check uses no patient data and establishes neither segmentation performance nor clinical validity. A native Lens interaction recording remains outstanding.

What stands out

Persistent layer stack: re-shading changes a key, not the geometry, so toggling a structure does not re-upload its mesh
Display-only presentation: a structure can be popped out, enlarged and turned while picks and measurements still resolve in source-scan millimetres
Honest 'this': answers are scoped per question (HU profile with partial-volume share, left/right comparison) and unsupported questions are refused with the reason
One vocabulary for voice, assistant and gesture (isolate, cut, keep, cancel); anatomical plane names are deliberately refused rather than guessed
Bifurcation markers drawn through the lumen wall, verified by an actual SceneKit render

Current limits

Never run with a real camera, microphone or iPad: gaze and hand interaction are tested in software only
The gaze/hand framework is gated behind a compile flag and has not landed in the Integration branch
macOS only (the stage uses SceneKit and AppKit)
Self-tests prove behaviour, not usability: there is no recorded user session
Non-diagnostic; surfaces are display aids over segmentation masks

Render check: seen through an opaque wall, with an ordinary-marker control; probe 8/8 (2026-10-03)

Review updates & corrections

New since 17 Oct bio: look/pinch/manipulate/ask integration (Sept 29 – Oct 5), assistant 'this' tools, the voice/gesture vocabulary, and bifurcation markers
The bio's 'files staged, not committed' is superseded: the work is committed on branch claude/lens-dicom-full-speed at 41edc5114 and merged UNCOMMITTED into the live Main tree; it is not in Integration
iOS: the Lens is macOS-only; gating fixed 2026-10-05
The 18 October 2026 website demonstration uses actual outputs from the AssistantApp offline DICOM decoder on a newly generated synthetic CT phantom. The 176 × 176 × 72 volume contains six known HU values. Decoded voxel counts and means for all six classes, and the whole-volume mean, match the independently generated ground truth. The browser viewer selects retained axial, narrow-window, coronal MPR and 60 mm maximum-intensity-projection images. It does not run the decoder in the browser and does not capture the native Lens interface. This bounded engineering check uses no patient data and establishes neither segmentation performance nor clinical validity. A native Lens interaction recording remains outstanding.

Workflow walkthrough

A guide to the native workflow. These steps describe the app; only the recordings and images embedded on this page are available as website demonstrations.

Open a contrast CT study and let segmentation finish; structures appear as layers on the Lens stage
Select a structure, then pop it out to enlarge and turn it, and reset it back to anatomy
Ask the assistant 'what is the HU profile of this?' and read the per-question answer and refusals
Say or type 'isolate this', then 'cut it across' and 'keep the cut'
Run the vessel workup and view the pink bifurcation markers through the vessel wall

Planned captures · not yet available

Pop out, turn and ask about 'this' on a phantom volume
Bifurcation markers visible through a vessel wall