💻 computer science

Comp2Comp: Open-Source Software with FDA-Cleared Artificial Intelligence Algorithms for Computed Tomography Image Analysis

The authors present Comp2Comp, an open-source software package featuring the first FDA-cleared deep learning pipelines for opportunistic computed tomography analysis, specifically for abdominal aortic quantification and bone mineral density estimation, demonstrating clinical-grade accuracy and improved transparency for medical imaging.

Original authors: Adrit Rao, Malte Jensen, Andrea T. Fisher, Louis Blankemeier, Pauline Berens, Arash Fereydooni, Seth Lirette, Eren Alkan, Felipe C. Kitamura, Juan M. Zambrano Chaves, Eduardo Reis, Arjun Desai, Marc H

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adrit Rao, Malte Jensen, Andrea T. Fisher, Louis Blankemeier, Pauline Berens, Arash Fereydooni, Seth Lirette, Eren Alkan, Felipe C. Kitamura, Juan M. Zambrano Chaves, Eduardo Reis, Arjun Desai, Marc H. Willis, Jason Hom, Andrew Johnston, Leon Lenchik, Robert D. Boutin, Eduardo M. J. M. Farina, Augusto S. Serpa, Marcelo S. Takahashi, Jordan Perchik, Steven A. Rothenberg, Jamie L. Schroeder, Ross Filice, Leonardo K. Bittencourt, Hari Trivedi, Marly van Assen, John Mongan, Kimberly Kallianos, Oliver Aalami, Akshay S. Chaudhari

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

The "Hidden Treasure" in Your Medical Scans: An Easy Guide to Comp2Comp

Imagine you are a detective. Every day, you walk through a massive warehouse filled with thousands of old, dusty boxes (these are your medical CT scans). Your job is to find specific clues—like a tiny crack in a foundation or a small leak in a pipe—that could predict a disaster later on.

The problem? You only look inside the boxes that someone specifically asked you to check. If a patient gets a CT scan because they have a stomach ache, you look at the stomach. You might walk right past a "clue" about their bones or their heart because nobody told you to look there. This is called "opportunistic imaging"—finding extra health clues in scans that were originally taken for something else.

Comp2Comp is like a high-tech, automated magnifying glass that scans every single box in that warehouse, looking for two specific "hidden treasures" that doctors often miss: Aortic Aneurysms and Osteoporosis.

1. The Aorta Watchdog (AAQ)

The Problem: Your aorta is the "superhighway" of your body—the massive pipe that carries blood from your heart to the rest of you. Sometimes, this pipe gets a bulge, like a weak spot in a garden hose. If that bulge (an aneurysm) gets too big, it can burst. This is a life-threatening emergency.

The Comp2Comp Solution: Think of the AAQ module as a precision digital ruler. Usually, a doctor has to manually measure that "hose" on a screen. Comp2Comp does it automatically. It looks at the scan, finds the exact widest part of the pipe, and measures it with incredible accuracy (within about 1.5 millimeters). It’s like having a tiny, tireless robot measuring the diameter of every pipe in a city to make sure none are about to burst.

2. The Bone Density Scout (BMD)

The Problem: Your bones are like the wooden beams of a house. As we age, those beams can become "honeycombed" or brittle (osteoporosis). If they get too weak, the house collapses (a bone fracture). The gold standard for checking this is a special test called a DXA scan, but many people never get it because it requires a separate appointment and extra cost.

The Comp2Comp Solution: Think of the BMD module as a "bonus check." While the doctor is looking at your scan for your original reason, Comp2Comp "peeks" at your spine. It calculates how dense your bones are by looking at the "texture" of the bone in the image. It’s like a building inspector walking by a house and, just by looking at the siding, being able to tell if the wooden frame inside is strong or rotting. It flags people who might need extra care before they actually break a bone.

Why is this a big deal? (The "Open Source" Revolution)

In the past, most medical AI was like a "Black Box." A company would sell it to a hospital, but the doctors couldn't see how it worked, how it was built, or if it might make mistakes. It was a "trust me, I'm a professional" situation.

Comp2Comp changes the game in three ways:

The "Open Recipe" (Open Source): Instead of a secret recipe, the creators shared the entire cookbook. Anyone—researchers, hospitals, or students—can see exactly how the AI "thinks." This makes it much safer and more transparent.
The "FDA Seal of Approval": Even though it’s free and open for everyone to use, it isn't "amateur hour." It has been officially cleared by the FDA (the government's top health watchdog), meaning it has passed rigorous tests and is proven to be accurate enough for real doctors to use.
The "Free Gift": Because it is "Open Source" (specifically under an Apache License), hospitals can use it, change it, and share it without paying massive licensing fees. It’s like a high-end tool that is being given away for free to help everyone live longer, healthier lives.

Summary

Comp2Comp is a digital assistant that turns every routine medical scan into a comprehensive health checkup. It hunts for silent threats—like bulging pipes or weakening bones—and does so using transparent, high-quality technology that anyone in the medical community can use to save lives.

Technical Summary: Comp2Comp Open-Source FDA-Cleared AI for CT Analysis

Problem: The "Transparency Gap" in Medical AI

While Artificial Intelligence (AI) enables "opportunistic imaging"—extracting valuable biomarkers (like bone density or aortic size) from existing CT scans without extra radiation or cost—the field faces two major hurdles:

Lack of Rigorous Validation in Open-Source Tools: Many free, open-source image analysis tools lack the formal clinical validation required for medical decision-making.
Lack of Transparency in Commercial Tools: Most FDA-cleared opportunistic analysis software is proprietary (closed-source). This creates a "transparency gap" where clinicians cannot audit the code, verify performance on their specific patient populations, or understand the underlying logic, leading to potential deployment failures.

Methodology: The Comp2Comp Platform

The authors developed Comp2Comp, a software framework providing two distinct deep-learning pipelines for opportunistic analysis. Both modules were developed using the nnU-Net framework, a self-configuring deep learning architecture for medical image segmentation.

1. Abdominal Aortic Quantification (AAQ):

Objective: Automatically measure the maximum diameter of the abdominal aorta to detect Abdominal Aortic Aneurysms (AAA).
Pipeline:
- Preprocessing: Filters for axial orientation and crops the volume to the L1–L5 lumbar region using TotalSegmentator.
- Segmentation: A 3D nnU-Net segments the aorta. To improve accuracy for aneurysms, the training set was enriched by upsampling a cohort of complex, aneurysmal cases from Stanford.
- Measurement: The algorithm fits an ellipse to each axial slice and selects the largest minor axis to represent the true diameter, mitigating errors caused by oblique vessel orientation.
Validation: Compared against the average of three board-certified radiologists across 258 scans from four international institutions.

2. Bone Mineral Density (BMD) Estimation:

Objective: Estimate vertebral trabecular bone density to identify osteopenia/osteoporosis risk.
Pipeline:
- Segmentation: Segments L1–L4 vertebrae.
- Calibration: To ensure accuracy across different scanners, the pipeline performs a two-point calibration using segmented visceral adipose tissue (VAT) and an air ROI to normalize Hounsfield Units (HU).
- Classification: A binary classifier compares the recalibrated HU score against a threshold (300 HU) to predict a DXA-equivalent T-score (Normal vs. Low).
Validation: Compared against the gold standard (DXA scans) for 371 patients across four institutions.

Key Contributions

First of its kind: The paper reports two of the first fully open-sourced, FDA 510(k)-cleared deep learning pipelines.
Full Transparency: Unlike commercial competitors, the authors released the code, model weights, and all pivotal statistical data submitted to the FDA.
Permissive Licensing: Released under the Apache License 2.0, allowing for commercial and non-commercial use, modification, and distribution.

Results

AAQ Performance:
- Mean Absolute Error (MAE): 1.57 mm (well below the 2.0 mm clinical threshold).
- Correlation: High agreement with radiologists (ICC of 0.983).
- Note: Accuracy was slightly lower in patients with aortic endografts and in scans with very low kVp (<100).
BMD Performance:
- Sensitivity: 81.0% (Threshold: >70%).
- Specificity: 78.4% (Threshold: >70%).
- Correlation: Pearson $r = 0.791$ for continuous density scores against DXA T-scores (Threshold: $>0.70$ ).

Significance and Clinical Impact

Comp2Comp bridges the gap between high-performance clinical tools and open science.

For Clinicians: It provides a validated tool to catch incidental findings (AAA and osteoporosis) that are often underreported in standard radiology workflows.
For Researchers: It provides "best-in-class" methods and a transparent blueprint for how to navigate the FDA clearance process.
For Healthcare Systems: By being open-source, it reduces the administrative and financial burden of "clinical pilots" for proprietary software, allowing hospitals to validate the tool locally before full-scale adoption.