Who actually made this 3D human body? Following the trail led to Japanese radio-wave research

In Part 1 I wrote about finding a 3D anatomy viewer on X and publishing a Japanese version of it.

Show only the bones, pick out muscles and organs, click on whatever catches your eye. Open a browser and spin it around.

The more I poked at it, the more a different question took over.

Hold on. Who made the underlying body data, and how?

I assumed somebody had opened Blender as a hobby and patiently carved a human being. Not even close. And the starting point wasn't anatomy at all.

What BodyParts3D is

In one line: a three-dimensional anatomy dictionary, where each part of the body is recorded not just as a name but as a 3D shape.

Look up "femur" in an ordinary dictionary and you get words. In BodyParts3D you also get the shape of the femur, where it sits in the body, and how it relates to the bones and muscles around it, as 3D data.

It was built and published by a team centred on the Database Center for Life Science (DBCLS), which has spent years organising life-science data from universities and research institutes. The project started around 2007; the first BodyParts3D came out in 2009.

That would be the end of the story — a research institute made anatomy teaching data — except there is another layer underneath.

It started with "how do radio waves get absorbed by the body?"

Mobile phones and wireless devices emit radio waves. Where in the human body do those waves get absorbed, and how much?

You can't answer that by running experiments on real people over and over under whatever conditions you like. What you need is a human body built inside a computer.

Researchers at the National Institute of Information and Communications Technology (NICT), Kitasato University, Keio University, Tokyo Metropolitan University and elsewhere developed numerical human models from whole-body MRI scans of Japanese adults, in order to calculate radio-wave absorption. The male model is called TARO, the female one HANAKO.

TARO was made by taking whole-body MRI scans of a volunteer close to the average height and weight of an adult Japanese male, then dividing the images into small cubes 2 mm on each side. Those cubes are called voxels: think of a pixel made solid.

The whole body is an enormous number of voxels, and each one carries tissue information: bone, muscle, fat, liver, and so on.

Today you'd assume AI did the extraction. At the time there was no automatic classification accurate enough, so medical staff went through the MRI images and classified the tissue by hand. The underlying research reports that identifying the tissue alone took more than three years.

Three years. Voxels for an entire body, looked at by people, sorted by people.

So TARO is not a 3D character. It is a research-grade body model, built from the whole-body MRI of a real person, with tissue identified by medical staff over a long stretch of time.

But TARO was not an anatomy dictionary

TARO's purpose was never to teach anatomy in detail. It was to calculate how radio waves are absorbed by the body.

So while the body was classified into about 50 tissues and organs, that is coarse if you look at it as teaching material.

There may be a category called "muscle", but you can't pick out the gluteus maximus, the gluteus medius, the quadriceps or the tibialis anterior and look at them individually. Fine for radio-wave calculations; not enough for "show me just this one muscle."

So a project began to take TARO — the whole-body foundation — and divide the body into finer anatomical parts. That project is BodyParts3D.

Subdividing the MRI doesn't finish the job

This is the part I found most interesting.

The original BodyParts3D paper describes the production process in three broad stages. The first stage added new anatomical divisions to TARO's data, splitting tissue that had been lumped together into finer organs, muscles and bones.

But an MRI does not show every part of the body crisply.

Thin nerves and vessels, small structures, the boundaries between adjacent muscles: all hard to make out. And some shapes simply cannot be represented at 2 mm spacing.

So in the next stage, 3D artists used dedicated software to clean up unclear outlines and fill in missing structures. They worked from Gray's Anatomy and other anatomy texts, brain atlases, clinical images, medical illustrations and scanned anatomical models.

Which means BodyParts3D is not MRI turned into 3D at the push of a button. What's inside is a mix of three things.

• Shapes taken from the MRI
• Shapes where a person cleaned up the outline, based on the MRI
• Shapes filled in or created by reference to anatomical sources

Put another way, it is closer to a blueprint of a standard human body: MRI as the foundation, with medical knowledge and human hands raising the level of finish.

Teaching the computer what each part is

Making the shapes finer still leaves you with a pile of 3D objects. To call it a database, the computer has to know what each lump actually is.

So in BodyParts3D, every part is mapped to an anatomical system called FMA (Foundational Model of Anatomy).

FMA is an anatomy ontology developed at the University of Washington, designed so that computers can work with the structure of the human body. "Ontology" sounds intimidating; here you can think of it as a very large classification table of anatomical names and relationships.

It teaches a computer relationships like these.

• A femur is a kind of bone
• The head of the femur is a part of the femur
• There is a left femur and a right femur
• Put the left and right femur together and you get the higher-level concept "femur"

BodyParts3D assigns each 3D shape an FMA identifier. Using "is a kind of" and "is a part of", larger anatomical concepts can then be assembled from smaller parts.

That is why BodyParts3D is a database rather than just human-body CG.

Why "is a kind of" and "is a part of" are kept separate

They look similar but do different jobs. "Is a kind of" is classification: follow the femur upward and you reach bone, then organ, and so on. "Is a part of" is assembly: collect the head and neck of the femur and you get the femur.

With both relationships in place, a request like "show every bone in the lower limb" can be resolved by the computer walking the graph, instead of a person clicking parts one at a time. That is the decisive difference from shipping the 3D shapes alone.

Who made it

BodyParts3D was a collaboration between people with different roles.

According to the official information, Kousaku Okubo, a physician, handled the project concept and design and the mapping of anatomical concepts. Kaori Fujieda and Shio Imai generated the part data; Nobutaka Mitsuhashi and Kaori Fujieda handled DICOM image processing.

Takuro Tamura, Nobutaka Mitsuhashi, Isamu Muto and Bits Co., Ltd. among others worked on system design and implementation, with DBCLS and NBDC maintaining it. The funding sits behind projects of MEXT and the Japan Science and Technology Agency (JST).

The work was published in 2009 in Nucleic Acids Research, a peer-reviewed journal in the life sciences.

So this is a very different animal from a free 3D asset of unknown origin. Who made it, from what, and for what purpose is all public, and the paper and the dataset are still there.

So how correct is it?

At this point it is tempting to think: research institutes, built from MRI, so it must be accurate. This is where it pays to slow down.

Here are the reasons to trust it, and the limits the project itself acknowledges.

That the official site is this candid about the limits raises my opinion of it, personally.

The other thing that matters: TARO was not made by averaging data from many people. It is based on the MRI of one particular Japanese man, close to average build.

Human bodies differ. By sex, age and build, and there are anatomical variations in vessels, nerves and muscles.

So BodyParts3D is not "a human body that matches every human". Useful as an anatomical map, but not a digital copy of any one person's body. That's the closest reading.

2013, and BodyParts3D 4.0

BodyParts3D has been updated since Release 1.0 in 2009.

• Release 1.0: February 2009
• Release 2.0: April 2010
• Release 3.0: June 2011
• Release 4.0: May 2013

Release 4.0 adjusted skeletal positions and added vessels inside the lungs and liver, cerebral arteries, arteries and veins throughout the body, the interior of the eye, and the pharynx and larynx.

The official archive still offers the BodyParts3D 4.0 polygon data in OBJ format, along with tables recording anatomical names and the relationships between parts. That published data is what gets reused as the foundation of body-viewing services, research and 3D models.

And then, a body viewer that runs in a browser

The 3D anatomy viewer I localised in Part 1, Human Atlas, stands on this published data.

Its author, ashemag, took the official BodyParts3D 4.0 distribution and rebuilt it as a modern web app running on React and Three.js. Not BodyParts3D itself: what they built is the part that displays 2,234 body parts smoothly in a browser and lets you search, select and pull them apart.

To do that, the original shapes were made lighter for the web (the repository says structures are simplified within a 0.2% error tolerance) and drawn in batches. The 15 systems shown on screen are a grouping organised for this app.

For what it's worth, the app code is MIT licensed and the anatomy data is CC BY with attribution required, and both are stated clearly. That is exactly why someone else, as in Part 1, can take it, change it and publish it. That's part of "publishing" too.

The interface looks current, and this is what's underneath it.

1. Scan the whole body of a Japanese man with MRI
2. Medical staff spend more than three years classifying tissue
3. Build TARO, a body model for radio-wave research
4. Divide TARO into finer anatomical parts
5. Fill in unclear shapes from anatomy texts and medical references
6. Link each part to an FMA anatomical concept
7. Publish the data for free
8. Another developer rebuilds it as an app that runs in a modern browser
9. Someone who saw that makes it readable in their own language

Laid out like that, the answer to "how on earth did someone build this?" was not a lone genius working fast. Researchers, medical staff, 3D artists and software developers built it over a long time, and another developer could build on it because the data was published. Near the end of the line, having only done the localisation, I get to ride along too.

That is what makes open data and open source fun.

What is genuinely impressive about BodyParts3D

Honestly, if you ask whether the BodyParts3D models are more beautiful than today's paid anatomy apps, not necessarily.

The data is old, the details contain errors and gaps, and it is based on one particular man.

What is impressive isn't that you can look at the shape of a body. It is that every part of the body was turned into data with shared coordinates and shared names, and published in a form anyone can reuse.

Normally, building this much body data from nothing needs MRI facilities, a subject, medical knowledge, anatomical references, 3D production skills, database design and web development. Not a scale one person can reach.

Because BodyParts3D is public, an individual developer can start from the top of that pile. Something you could never build alone comes within reach by following published research data.

What surprised me about Human Atlas wasn't only the flashy 3D. It was the twenty-plus years of research and patient human work behind it, and the culture of publishing it so the next person can use it.

If any of this interests you, start by looking at the official download page. What is there, in what format, under what conditions. Even just looking at that tells you how much work sits under the phrase "published for free".

And if you want to actually touch it, Part 1 is the place. That's the one about localising the 3D viewer I found on X.

→ Part 1: I found a 3D human body viewer built with GPT-6 Astra, so I translated it into Japanese

To try it yourself → human-atlas.pages.dev

Sources

• Part 1: I found a 3D human body viewer built with GPT-6 Astra
• BodyParts3D database description
• BodyParts3D official information: team and production process
• BodyParts3D Release 4.0 release notes
• BodyParts3D download page
• Mitsuhashi et al. BodyParts3D: 3D structure database for anatomical concepts
• Nagaoka et al. on the development of TARO and HANAKO
• Human Atlas repository (ashemag)
• The Japanese version of Human Atlas (Cloudflare Pages)

---

The production team, release history and data limitations are summarised from the official sites and the original paper as of 6 September 2026. Before reusing the data, always check the official licence statement.