Contents

A voxel illustration of Gekun checking a 3D anatomy app against reference materials

In part 1, I translated a 3D anatomy viewer I found on X into Japanese.

Following its source data led me to research in Japan, and then to Z-Anatomy on the other side of the world.

Meanwhile, I kept working on the app.

Looking back through the development notes, even I was surprised by how much had changed. Here's where things stand.

First, it has a new name

I originally published it as a Japanese version of Human Atlas, keeping the original project's name.

As I added features, “just a translation” no longer felt accurate. So I renamed it JINTAI 3D and moved it to a new address.

→ jintai-3d.pages.dev

The old human-atlas.pages.dev address redirects to the new one.

There are now two models

This is the biggest change.

I added the Z-Anatomy model discussed in part 3: bones, muscles and connective tissue, plus 178 peripheral nerve parts in the arms and legs, for a total of 1,138 parts.

The original BodyParts3D is useful for exploring the whole body and its organs. Z-Anatomy is useful for seeing the relationships between bones, muscles and nerves.

Rather than choosing one, I made both available at launch. The opening screen offers “Whole body and organs” on the left and “Bones, muscles and peripheral nerves” on the right, with previews rendered from the actual 3D data.

The licenses also differ: BodyParts3D uses CC BY 4.0 and Z-Anatomy uses CC BY-SA 4.0. The app includes a comparison of their coverage, uses and licenses.

I've improved how it feels to use

These changes are less dramatic, but there are a lot of them.

  • Tapping a part used to open its explanation immediately, covering the model. Now it first shows buttons for the explanation, hiding the part and clearing the selection.
  • Selected parts were hard to spot. They now get blue or rose highlighting, depending on their color, with a gentle pulse that stops when reduced motion is enabled.
  • I wanted to zoom in on a selected part and rotate it. A new Focus action smoothly moves in, then smoothly returns.
  • Hidden parts can be restored with Undo and Restore all.
  • Multiple parts can be selected by tapping them.
  • Removing the floor and pedestal lets you rotate underneath the feet.
  • Shortcuts move the view to the shoulder, lower back, knee or foot.

My work involves feet, so not being able to look up at the soles was particularly frustrating. Removing the floor sounds minor, but it might be my favorite improvement.

I also found a face-orientation mismatch in mirrored parts that caused odd shading on structures such as the left plantar aponeurosis. I corrected 547 parts.

Making it playable: Anatomy Challenge

I wanted to see whether the app could help with learning, beyond browsing it as a dictionary. So I built some quizzes.

  • Touch Challenge: find a named superficial muscle on the 3D model, with five questions against the clock.
  • What is this part?: name the highlighted muscle.
  • Guess the shape: identify a muscle from its isolated shape.

They start with a 3–2–1–GO countdown, and a large circle appears for a correct answer. Results can be shared.

One thing I learned while making them: knowing a name and being able to point to it in 3D are very different skills. Even I get more wrong than I expected.

Making it move: experimental joint motion

This is still explicitly labeled an experimental feature in the app.

Bone models can now demonstrate basic movements of the elbow, wrist, knee, ankle, hip, shoulder and neck.

This turned out to be difficult. Simply rotating bones around a single axis makes them intersect. Real joint movement involves rolling and gliding.

For the elbow and knee, I've been trying a provisional arc-based approximation of that rolling and gliding. For the knee, I also added a mode that plays back published results from the OpenSim JAM research model, using Lenhart 2015 data for approximately 0–90° of passive right-knee flexion.

None of these should be described as an exact reproduction of real joint movement. The app says so, and I intend to keep saying so.

What takes most of my time now

So far, this probably sounds like a steady stream of visual changes and new features.

But most of my time over the last few days has gone into something that doesn't show up on screen.

Checking whether the names and explanations are actually correct.

JINTAI 3D has explanations for general readers and clinical readers for individual structures. I'm going through them one by one, opening the sources and checking them again. There are roughly 3,900 research tasks across nerves, muscles, connective tissue, the skeleton, organs and reviews of existing explanations.

Here's the division of work.

Role Who handles it
Overall planning and independent review Codex
Work design, instructions and acceptance decisions Claude
Research and implementation Cursor
Final check before publication Me, with specialist input when needed

The key is to have different AIs do the research and the checking. The same writer doesn't get to approve its own work.

Looking closer turned up quite a lot

More than I expected, in fact.

  • Parts displayed as skeletal structures included muscles such as tibialis anterior. This has been corrected.
  • Some heart-valve elements included leaflets from a different valve.
  • A brain ventricle was classified under the cardiac system. My interpretation is that the shared English word “ventricle” caused an automated mix-up.
  • There were 29 recorded cases of explanations being linked to papers unrelated to the intended anatomy topic.

I'm not trying to dismiss the work of the people who created the original data. As I wrote in part 3, their decision to publish it is what makes this project possible. But if I'm going to display it in my own app, checking and correcting it is my responsibility.

I don't simply trust the AI doing the checking, either

Another interesting discovery: the checking process can have problems of its own.

In the skeleton research, about 40% of the work submitted for review was judged to need corrections. We paused new assignments, revised the research instructions and then resumed.

There was also a case where the supervising AI counted files that were still being written and reported more accepted results than actually existed. We caught and corrected that, and changed the process to check that writing had finished before counting files.

AI can work fast. But “I'm finished” and actually being finished are different things. Building a process that closes that gap has become the biggest part of my work.

What's next

I'm thinking in three stages.

  1. An anatomy dictionary: find a structure, understand its position and relationships, and check basic information against sources. I also plan to add a 3D layer for meridians and acupuncture points.
  2. Anatomy learning: use quizzes and joint motion to help with understanding and retention.
  3. Uses beyond healthcare: help people interested in strength training, yoga and understanding their own movement.

I'm still partway through the first stage. Research results will be incorporated into the app after independent review and my own check of the content.

There may be fewer visually dramatic updates for a while. What I most want now is to move toward a state where people can trust what the app says.

Before using the app

JINTAI 3D is a learning viewer, not a tool for diagnosis or surgery. Some names remain in English, and expert review of anatomical accuracy is not yet complete. Joint motion is experimental and does not precisely reproduce real joint movement.

This series

Counts and features reflect development records and the update history inside JINTAI 3D as of September 10, 2026. The published simulation results used for joint motion come from the OpenSim JAM passive right-knee flexion example.