🤖 AI

The stochastic digital human is now enrolling for in silico imaging trials -- Methods and tools for generating digital cohorts

This paper reviews the latest methods and tools for generating stochastic digital human models to facilitate *in silico* imaging trials, covering model classification, generation techniques for healthy and diseased states, and the trade-offs involved in sampling digital cohorts to minimize study bias.

Original authors: A Badano, M Lago, E Sizikova, JG Delfino, S Guan, MA Anastasio, B Sahiner

Published 2026-05-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: A Badano, M Lago, E Sizikova, JG Delfino, S Guan, MA Anastasio, B Sahiner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor or a medical device maker. You have a new X-ray machine or a new way to scan the body, and you need to prove it works before you can sell it to hospitals. Usually, you have to run a Clinical Trial. This means finding hundreds of real people, scanning them, and seeing if the machine finds diseases correctly.

This paper argues that this process is slow, expensive, and sometimes risky for the people involved. Instead, the authors propose a "video game" version of this trial. They want to build a virtual world filled with digital humans to test the machines.

Here is a breakdown of their ideas, using simple analogies:

1. The Goal: The "Virtual Stadium"

Think of a real clinical trial like a massive sports tournament. You need real athletes (patients) to play the game. But finding the right players is hard; some might get hurt, some might drop out, and you might not have enough players from rare backgrounds (like very tall people or people with a specific rare disease).

The authors want to build a Virtual Stadium (called an in silico trial). In this stadium, the "athletes" are Stochastic Digital Humans.

"Stochastic" is a fancy word for "random." It means these digital humans aren't all identical clones. They are generated with random variations in size, shape, and tissue density, just like real people.
The goal is to create a crowd of thousands of these digital people so you can test your medical device on a huge, diverse group without ever asking a real person to step into the scanner.

2. The Cast of Characters: How Do We Build Them?

The paper explains that there are different ways to build these digital people, similar to how you might build a character in a video game.

A. The "Photo-Real" Avatars (Individual Models)

Personalized Models: Imagine taking a specific person's MRI scan and turning it into a 3D digital twin. This is great for planning surgery for that one person, but it's not useful for testing a machine on a whole population because you'd need to scan millions of real people first.
Family Models: Imagine a small "family" of digital people (a dad, a mom, and two kids) based on real scans. This is better, but still too small to represent the whole world.

B. The "Generators" (Population Models)
To get a huge crowd, you need a machine that can create people from scratch. The paper discusses two main ways to do this:

The "Mold" Method (Image-Based): You take a pile of real medical images (like thousands of X-rays) and use math to stretch, shrink, and mix them to create new ones.
- The Catch: If your real X-rays are blurry or have noise, your digital people will inherit those flaws. It's like trying to bake a perfect cake using a recipe that has smudged ink on it.
The "Blueprint" Method (Knowledge-Based): Instead of copying real photos, you write a set of rules (a blueprint) based on biology. You tell the computer: "Make a breast that is 50% fat and 50% glandular tissue, with blood vessels branching out like a tree."
- The Benefit: You can generate millions of unique people this way, including rare types that might not exist in your photo database.
- The Catch: If your rules are wrong, you might create a digital human with a liver made of sponge instead of tissue.

3. Adding the "Bad Guys" (Modeling Disease)

A medical trial isn't just about healthy people; it's about finding the sick ones. How do you put a tumor into a digital human?

The "Sticker" Method (Image-Based): You take a real picture of a tumor and paste it onto a digital body.
- Problem: It looks a bit fake, like a sticker on a painting. It doesn't interact with the body around it.
The "Growth" Method (Knowledge-Based): You program the tumor to "grow" according to biological rules. If the surrounding tissue is stiff, the tumor grows a certain way; if it's soft, it grows differently.
- Benefit: This creates a much more realistic simulation of how a disease actually behaves.

4. The "Photo Editor" Trick (Augmentation)

Sometimes, you don't have enough digital people. So, you take the ones you have and "edit" them.

You can rotate them, change the brightness, or add "static" (noise) to the image, just like using filters on a photo app.
The Warning: The authors warn that if you edit too much, you might create something impossible. For example, if you accidentally make a bone look softer than skin, your test results will be garbage. You have to be careful not to break the laws of physics while trying to make more data.

5. The Most Important Lesson: Don't Cheat the Sample

This is the paper's biggest warning. Imagine you are testing a new shoe.

If you only test the shoes on runners (a specific group), you might think the shoes are great.
But if you actually sell them to everyone (including people who walk, dance, or have flat feet), the shoes might fail.

In the digital world, it is very easy to accidentally pick a "biased" group of digital humans.

The Trap: If you randomly pick digital humans without thinking, you might end up with a group that is mostly average-sized people. You might miss the "outliers" (very large or very small people).
The Result: The paper shows that if you pick the wrong group, your test results can be completely wrong. You might think a machine is 90% accurate when it's actually only 77% accurate for the real world.

Summary

The paper is a guidebook for building a virtual testing ground for medical devices.

Why? Real trials are slow, expensive, and hard to organize.
How? We build "Stochastic Digital Humans" using either real data (photos) or biological rules (blueprints).
The Catch: We must be very careful to make sure our virtual crowd looks exactly like the real world. If we pick the wrong "digital people" to test on, we might get the wrong answer, and that could be dangerous for real patients later.

The authors conclude that while we have the tools to build these digital crowds, we still need to figure out exactly how to make sure they are realistic enough to trust with people's lives.

Technical Summary: Methods and Tools for Generating Digital Cohorts for In Silico Imaging Trials

Problem Statement
Randomized clinical trials (RCTs) are the standard for evaluating medical interventions but suffer from limitations including safety risks, difficulties in patient recruitment and retention, under-representation of rare subpopulations, and high resource costs. While in silico (computational) trials offer a pathway to overcome these barriers by simulating device performance entirely in the digital world, they require realistic digital representations of human subjects. A critical gap exists in the methodology for generating large, stochastic digital cohorts that accurately reflect the variability of physical-world patient populations, including both healthy and diseased states. Without these models, in silico trials cannot achieve the statistical power or representativeness required for regulatory evaluation.

Methodology and Framework
The paper proposes a classification framework for "stochastic digital humans" and reviews current methodologies for their generation. The authors distinguish between:

Individual Models (Non-stochastic): Deterministic replicas of specific physical exemplars (e.g., "digital twins" or "family models" like the Virtual Family). These are useful for specific tasks but lack the statistical variability required for population-level trials.
Population Models (Stochastic): Models designed to sample from distributions representing human variability. These are further categorized into:
- Image-based Models: Derived from acquired medical images.
  - Parametric: Use mathematical equations (e.g., splines, NURBS) and deformation techniques (morphing, registration) based on limited patient datasets (e.g., XCAT phantom). These are limited by the size of the source dataset and the "null space" of the imaging system (information loss during acquisition).
  - Generative: Utilize neural networks (e.g., GANs, Diffusion models) to synthesize populations from image data. While promising for high-resolution synthesis, they face challenges such as mode collapse, reliance on degraded image data rather than true object properties, and the "chicken-and-egg" problem of needing high-quality object data to train models that reconstruct objects from noisy images.
- Knowledge-based (Procedural) Models: Constructed by sampling parameters ( $\theta_n$ ) and expansion functions ( $\phi_n$ ) from distributions derived from physical or biological measurements. These allow for unlimited random realizations and the inclusion of rare cases (e.g., the VICTRE breast model using Voronoi segmentation and recursive branching). However, they are computationally intensive and require extensive prior knowledge to define parameters accurately.

Modeling Disease
The paper outlines two primary approaches for incorporating disease into digital cohorts:

Image-based: Extracting lesion characteristics directly from patient data. This is limited by resolution and the availability of diverse lesion morphologies.
Knowledge-based: Generating lesions using procedural algorithms (e.g., simulating tumor growth based on tissue stiffness and elasticity). This allows for unlimited lesion shapes and variable resolutions but often relies on simplifying assumptions, such as treating disease as an additive component independent of the underlying anatomy, which may not fully capture architectural distortions caused by growing tumors.

Augmentation Methods
The authors review data augmentation techniques (geometric transformations, intensity operations, noise injection, and object combination) used to expand datasets. While practical and fast, the paper notes significant limitations: augmented data may be biologically or physically implausible (e.g., incorrect Hounsfield units) and often fails to generate truly independent new representations of the population, merely creating new combinations of existing inputs.

Key Results and Analysis
The paper presents a theoretical analysis of sampling strategies for digital cohorts using a case study comparing Digital Mammography (DM) and Digital Breast Tomosynthesis (DBT). By simulating four different sampling approaches (uniform, matched, simpler, and narrow) against a bimodal target population distribution, the authors demonstrate that the choice of sampling strategy significantly impacts performance assessment.

Result: The difference in Area Under the Curve (AUC) between the two imaging modalities varied from 0.01 to 0.11 depending on the sampling strategy.
Implication: Selecting a specific patient distribution for a digital cohort introduces potential study bias. If the digital distribution ( $f_d$ ) does not closely approximate the intended physical-world distribution ( $f_i$ ), the trial results may be misleading.

Significance and Claims
The paper positions itself as a review and classification of the current state of the art rather than a presentation of new experimental results. Its primary contributions are:

Terminology Standardization: Proposing the term "stochastic digital human" to distinguish population-level models used for regulatory trials from "digital twins" (individual-specific models).
Methodological Survey: Providing a comprehensive taxonomy of image-based, knowledge-based, and generative approaches, highlighting their respective trade-offs regarding realism, variability, and computational cost.
Bias Awareness: Emphasizing that the generation of digital cohorts is not merely a technical exercise but a statistical one. The paper argues that careful study design is required to minimize the statistical distance between digital and physical populations to ensure valid regulatory conclusions.

The authors conclude that while in silico trials hold significant promise for reducing time, cost, and risk in medical device evaluation, the field must advance in evaluating the quality of digital human models. Future work must rigorously define the patient distributions these models can capture to prevent misuse and ensure patient safety, particularly for rare diseases and underrepresented subgroups.