🔬 physics

OpenPros: A Large-Scale Dataset for Limited View Prostate Ultrasound Computed Tomography

This paper introduces OPENPROS, the first large-scale benchmark dataset comprising over 280,000 realistic 2D speed-of-sound maps and corresponding ultrasound waveforms derived from clinical and ex vivo prostate models, designed to advance and rigorously evaluate machine learning methods for limited-view prostate ultrasound computed tomography.

Original authors: Hanchen Wang, Yixuan Wu, Yinan Feng, Peng Jin, Luoyuan Zhang, Shihang Feng, James Wiskin, Baris Turkbey, Peter A. Pinto, Bradford J. Wood, Songting Luo, Yinpeng Chen, Emad Boctor, Youzuo Lin

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanchen Wang, Yixuan Wu, Yinan Feng, Peng Jin, Luoyuan Zhang, Shihang Feng, James Wiskin, Baris Turkbey, Peter A. Pinto, Bradford J. Wood, Songting Luo, Yinpeng Chen, Emad Boctor, Youzuo Lin

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 Big Problem: Finding a Needle in a Haystack (But the Haystack is Your Body)

Imagine you are trying to find a small, hidden object inside a complex, crowded room. You can only stand in two specific corners of the room (the front and the side) to look inside. You cannot walk around the room to see the back. This is the challenge doctors face when trying to image the prostate (a small gland in men) using ultrasound.

The Goal: Detect prostate cancer early.
The Current Tool: Standard ultrasound is like looking through a foggy window. It's cheap and fast, but it often misses tumors, especially those hidden in the front of the prostate or behind bones.
The Better Tool: A new technology called Ultrasound Computed Tomography (USCT). Instead of just taking a picture, it measures how fast sound waves travel through tissue. Different tissues (like healthy cells vs. tumors) slow down sound waves differently. If we can map these speeds perfectly, we can see tumors clearly.

The Catch: Because of the body's anatomy (bones and other organs), we can't surround the prostate with sensors. We are stuck with a "limited view," which makes the math to reconstruct the image incredibly difficult and slow.

The Solution: OPENPROS (The Ultimate Training Gym)

To fix this, the researchers created OPENPROS. Think of this as a massive, high-tech flight simulator for doctors and AI.

Before a pilot can fly a real plane, they need thousands of hours of practice in a simulator that mimics real storms, engine failures, and bad weather. Similarly, before an AI can help doctors diagnose cancer, it needs to practice on millions of "fake" but realistic cases.

What is in the OPENPROS dataset?

The "Ground Truth" (The Answer Key): They took real MRI and CT scans from patients and actual prostate specimens from a lab. They used these to build a perfect 3D digital map of what a prostate should look like, including where the bones are and how different tissues sound.
The "Simulated Data" (The Test Questions): They used super-computers to run a physics simulation. They "fired" virtual sound waves at these digital prostates and recorded what the waves looked like when they bounced back.
The Scale: They created 280,000 pairs of these "questions" (sound waves) and "answers" (the perfect map). This is a huge library for AI to learn from.

How They Tested the AI (The Race)

The researchers didn't just build the dataset; they used it to race two types of "drivers" to see who could solve the puzzle best:

The Old School Driver (Physics-Based Methods): These are traditional math algorithms that try to solve the sound wave equations step-by-step.
- Result: They are accurate but painfully slow. It takes them hours to process one image. They also get confused easily when the view is limited, resulting in blurry pictures.
The New Driver (Deep Learning/AI): These are neural networks (like the brains behind self-driving cars) trained on the 280,000 examples.
- Result: They are incredibly fast. They can process an image in milliseconds (faster than a blink). They produce much clearer pictures than the old methods.

The Good News and The Bad News

The Good News:
The AI is a game-changer for speed. It can turn a process that used to take hours into a process that takes a fraction of a second. It also sees the general shape of the prostate much better than the old methods.

The Bad News (The "Uncanny Valley" of Medical Imaging):
Even though the AI is fast and generally accurate, it still struggles with the fine details.

Analogy: Imagine the AI is a master painter. It can paint a beautiful landscape that looks exactly like the real thing from a distance. But if you zoom in with a magnifying glass, the tiny leaves on the trees and the texture of the bark are still a bit blurry.
In the paper, this means the AI can find the general location of a tumor, but it sometimes misses the tiny, sharp edges or small details inside the prostate. It also struggles when it sees a patient it has never "met" before (a new anatomy it wasn't trained on).

Why This Matters

The paper argues that we cannot fix these "blurry details" or make the AI smarter without a good training library. Before OPENPROS, researchers didn't have a large, realistic dataset specifically for this difficult "limited view" prostate problem.

By releasing OPENPROS to the public, the authors are handing the keys to the whole research community. They are saying: "Here is the simulator, here are the test questions, and here are the answers. Go ahead and build better AI drivers that can see the tiny details and handle new patients without getting confused."

Summary

The Problem: Seeing prostate cancer is hard because of body anatomy, and current tools are slow or blurry.
The Tool: OPENPROS, a massive dataset of 280,000 simulated ultrasound scans based on real human anatomy.
The Discovery: AI is much faster and clearer than old math methods, but it still needs to get better at seeing tiny details and handling new patients.
The Goal: To give researchers a standard playground to build the next generation of cancer-detecting tools.

Technical Summary: OPENPROS

Problem Statement

Prostate cancer remains a leading cause of cancer-related mortality, necessitating early detection. While multiparametric MRI (mpMRI) is the current gold standard, its high cost and limited accessibility hinder widespread adoption. Ultrasound Computed Tomography (USCT) offers a cost-effective alternative by reconstructing quantitative tissue parameters, such as the speed of sound (SOS), which serve as biomarkers for malignancy. However, prostate USCT faces significant clinical and technical barriers:

Limited-Angle Acquisition: Anatomical constraints restrict transducer placement to transrectal and transabdominal positions, creating sparse, angularly limited data that renders traditional physics-based inversion methods ill-posed, slow to converge, and prone to severe artifacts.
Tissue Complexity: The prostate environment involves high tissue heterogeneity and proximity to bony pelvic structures, invalidating simplified fluid assumptions used in other modalities like breast imaging.
Data Scarcity: The development of robust, generalizable reconstruction algorithms is severely hampered by the lack of large-scale, anatomically precise datasets that simulate realistic limited-angle conditions with full-waveform data. Existing datasets are either focused on breast imaging, lack anatomical realism, or do not address the specific challenges of the male pelvic region.

Methodology

The paper introduces OPENPROS, a large-scale benchmark dataset and simulation framework designed to bridge the gap between academic research and clinical deployment for limited-angle prostate USCT.

Dataset Construction:

Source Material: The dataset is derived from 4 real clinical MRI/CT scans and 62 ex vivo prostate specimens. Expert annotations were performed on T2-weighted MRI (organs), T1-weighted MRI (fat), and X-ray CT (bones).
Physical Measurements: Speed of sound (SOS) and attenuation were measured from ex vivo specimens using a QT scanner. Other organ SOS values were sourced from the ITIS foundation tissue database.
Phantom Generation: 3D digital prostate models were constructed by integrating these anatomical segmentations and measured SOS distributions. To mimic tissue heterogeneity, Gaussian distributions were applied to SOS values within tissue types.
2D Extraction: 280,000 paired 2D SOS phantoms and ultrasound waveforms were generated by slicing the 3D models under clinically realistic probe configurations (one transrectal, one abdominal) with random rotations (±45°) and perturbations.
Simulation: Wave propagation was simulated using a fourth-order spatial and second-order temporal Finite-Difference Time-Domain (FDTD) solver and Runge-Kutta acoustic wave solvers. The setup includes 20 sources and 322 receivers, capturing transmission and reflection across diverse acoustic paths. The simulation accounts for absorbing boundary conditions and uses a 1 MHz Ricker wavelet.

Benchmarking Framework:
The authors established a supervised learning task to approximate the inverse mapping from ultrasound pressure fields ( $p$ ) to SOS maps ( $c$ ). They compared:

Physics-based Baselines: Delay-and-Sum beamforming and a multi-stage, multi-frequency USCT inversion method.
Deep Learning Baselines: A CNN-based InversionNet and a Vision Transformer (ViT)-based variant (ViT-Inversion) designed to capture long-range spatial dependencies.

Key Contributions

OPENPROS Dataset: The first large-scale, publicly available benchmark specifically for limited-angle prostate USCT. It contains over 280,000 paired samples of realistic 2D SOS phantoms and corresponding full-waveform ultrasound data, derived from clinical and ex vivo sources.
Open-Source Simulation Tools: The release of advanced FDTD and Runge-Kutta acoustic wave solvers to ensure reproducibility and facilitate the development of new inversion methods.
Comprehensive Benchmarking: A systematic evaluation of physics-based and deep learning methods under realistic limited-angle conditions, establishing performance baselines for inference efficiency, reconstruction accuracy, and generalization.

Results

In-Distribution (ID) Performance:

Accuracy: Deep learning methods significantly outperformed traditional physics-based algorithms. ViT-Inversion achieved the best results with an RMSE of ~0.0205 and SSIM of ~0.9967, representing a 5–6× reduction in error compared to the physics-based baseline (RMSE ~0.16).
Efficiency: Data-driven methods demonstrated near-instantaneous inference (4.9ms for InversionNet, 8.9ms for ViT-Inversion) compared to the hours required for physics-based methods (4–24 hours per sample).
Qualitative Limitations: While deep learning models recovered global anatomical shapes and boundaries more clearly than physics-based methods, they failed to resolve fine internal structures and small lesions, often producing smoothed reconstructions.

Generalization and Robustness:

Patient-Level Generalization: When tested on entirely unseen patients (Out-of-Distribution), performance degraded significantly (errors increased by 3–5×), highlighting the difficulty of generalizing across different anatomies.
Intra-Patient Generalization: Models performed well when tested on unseen prostates from the same patient, suggesting intra-patient variability is less challenging than inter-patient variability.
Noise Robustness: ViT-Inversion demonstrated superior resilience to Gaussian noise compared to InversionNet.

Significance and Claims

The paper positions OPENPROS as a critical resource to motivate advances in physics-informed learning, foundation models for scientific imaging, and uncertainty-aware reconstruction.

The authors claim that while deep learning offers a pathway to rapid, clinically relevant SOS reconstruction, current methods still fail to deliver the high-resolution, clinically acceptable reconstructions required for real-world deployment. The dataset is intended to:

Enable fair and rigorous method comparisons.
Expose critical gaps in generalization and robustness that must be addressed before clinical translation.
Serve as a fast-prototyping benchmark for waveform-to-SOS reconstruction, with the expectation that successful approaches will eventually be validated in more complex 3D, multi-parameter models.

The work concludes that OPENPROS bridges the gap between academic ML research and clinical needs, but emphasizes that significant challenges remain in resolving fine anatomical details and achieving robust generalization across unseen anatomical variations.