⚡ electrical engineering

PSIRNet: Deep Learning-based Free-breathing Rapid Acquisition Late Enhancement Imaging

This paper introduces PSIRNet, a deep learning method that reconstructs diagnostic-quality free-breathing late gadolinium enhancement cardiac MRI images from a single two-heartbeat acquisition, achieving an 8- to 24-fold reduction in scan time while demonstrating superior or equivalent image quality compared to traditional motion-corrected methods.

Original authors: Arda Atalik, Hui Xue, Rhodri H. Davies, Thomas A. Treibel, Daniel K. Sodickson, Michael S. Hansen, Peter Kellman

Published 2026-04-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arda Atalik, Hui Xue, Rhodri H. Davies, Thomas A. Treibel, Daniel K. Sodickson, Michael S. Hansen, Peter Kellman

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 trying to take a perfect photograph of a moving object, like a hummingbird, but you can only take a blurry, noisy snapshot every time you press the shutter. To get a clear picture, the old way of doing things was to take that blurry snapshot 8 to 24 times, hoping that if you stacked them all together and smoothed out the noise, you'd eventually get a sharp image.

The Problem:
In the world of heart MRI scans (specifically looking for scar tissue called "Late Gadolinium Enhancement"), doctors have to do exactly this. They ask the patient to hold their breath and take many, many "snapshots" (heartbeats) to build a clear picture.

The Cost: This takes a long time (15+ minutes), makes the patient hold their breath repeatedly (which is hard for sick people), and clogs up the MRI machine schedule.
The Computer Work: After taking all those photos, a computer has to spend hours trying to align them and clean up the noise. It's like trying to assemble a puzzle where the pieces keep moving.

The Solution: PSIRNet
The researchers in this paper built a super-smart AI called PSIRNet. Think of PSIRNet as a Master Chef who has tasted millions of dishes.

The Training: They fed this AI a massive library of over 800,000 heart images from 55,000 different patients. It learned what a "perfect" heart scan looks like by studying the results of those long, slow, multi-snapshot scans.
The Magic Trick: Now, instead of asking the patient to hold their breath for 24 heartbeats, the AI only needs two heartbeats (one quick snapshot).
The Result: The AI looks at that single, blurry, noisy snapshot and instantly "imagines" what the perfect, clear image should look like based on everything it learned. It fills in the missing details and removes the noise better than the old method of stacking 24 photos.

Why is this a big deal? (The Analogies)

The Speed Boost:
- Old Way: It's like trying to cross a river by hopping on 24 floating logs, one by one, while the water rushes by.
- PSIRNet: It's like building a bridge in a split second. The scan time is reduced by 8 to 24 times. What used to take 15 minutes now takes a few seconds of scanning.
The "Breath-Hold" Relief:
- Old Way: You have to hold your breath for a long time, which is scary and difficult for patients with heart or lung issues.
- PSIRNet: The patient can just breathe normally. The AI is so good at "motion correction" (fixing the blur caused by breathing) that it doesn't care if the patient is moving slightly.
The Computer Speed:
- Old Way: The computer takes 5+ seconds to process each slice of the heart. If you have 20 slices, that's over a minute of waiting just for the computer to think.
- PSIRNet: The AI processes a slice in 0.1 seconds (100 milliseconds). It's like the difference between waiting for a slow dial-up internet connection and having 5G. The doctor can see the results almost instantly while the patient is still in the machine.

Did it work?
The researchers tested this on a huge group of patients and had two expert heart doctors (cardiologists) grade the pictures.

The Verdict: The doctors said the AI-generated pictures were just as good as, or even better than, the old slow method. They could clearly see the scar tissue and the heart boundaries.
The Variety: It worked for different types of heart scans (bright blood, dark blood, and scans for patients with pacemakers).

The Bottom Line
PSIRNet is like giving the MRI machine a "superpower." It allows doctors to get high-quality, life-saving diagnostic images in a fraction of the time, with the patient breathing comfortably. This means more patients can be scanned, the scans are more comfortable, and the results are available instantly.

Note: While the AI is amazing, the researchers are still doing final checks to make sure it works perfectly in real-world hospitals before it becomes the standard for everyone.

1. Problem Statement

Late Gadolinium Enhancement (LGE) cardiac MRI is the clinical gold standard for assessing myocardial viability, infarction, and scarring. However, current clinical workflows face significant bottlenecks:

Time-Consuming Acquisition: Conventional free-breathing LGE requires 8 to 24 motion-corrected (MOCO) signal averages to achieve sufficient Signal-to-Noise Ratio (SNR). Each average takes two heartbeats (one Inversion Recovery [IR] and one Proton Density [PD] image), leading to scan times of 15+ minutes per study.
Computational Burden: The conventional reconstruction pipeline involves non-rigid motion correction, registration, and averaging, which takes >5 seconds per slice on a CPU. This delays image availability and limits volumetric coverage.
Limitations of Existing DL Methods: Previous deep learning approaches for LGE have largely focused on breath-held, segmented acquisitions with modest acceleration factors and small, single-center datasets, lacking generalizability across different scanner vendors, field strengths, and LGE variants (e.g., dark blood, wideband).

Goal: Develop a method to generate diagnostic-quality Phase-Sensitive Inversion Recovery (PSIR) images from a single acquisition (2 heartbeats), eliminating the need for extensive averaging and reducing scan time by 8–24 fold.

2. Methodology

A. Dataset and Scope

Scale: The study utilized a massive retrospective dataset comprising 800,653 slices from 55,917 patients.
Diversity: Data was acquired between 2016–2024 across multiple institutions, 1.5T and 3T scanners, and various software versions.
Variants: The model was trained and tested on three distinct LGE protocols:
1. Bright Blood: Standard SSFP readout.
2. Dark Blood: Includes T2 preparation.
3. Wideband: Designed for patients with implanted cardiac devices (requires high bandwidth).
Data Split: Strictly patient-level separation to prevent leakage:
- Training: 640,000 slices (42,822 patients).
- Validation: 80,014 slices (4,952 patients).
- External Testing: 80,639 slices (8,143 patients) from different institutions.

B. Model Architecture: PSIRNet

PSIRNet is an end-to-end, physics-guided deep learning network with 845 million parameters. It operates on a single interleaved IR/PD acquisition (the first heartbeat pair) rather than averaging multiple acquisitions.

Input: Noise-prewhitened, undersampled k-space data for IR and PD, plus initial coil sensitivity maps.
Core Mechanism: The network unrolls 12 cascaded Landweber iterations.
1. Sensitivity Map Refinement (SMR): A 4-level U-Net refines coil sensitivity maps.
2. Joint IR-PD Refinement: A 4-level U-Net (96 first-level channels) performs joint refinement of the IR and PD images.
3. Data Consistency: Each iteration includes a learned data-consistency step with adaptive step sizes ( $\lambda$ ) to ensure the reconstruction adheres to the acquired k-space data.
Output Generation: After the final iteration, the network removes the background phase of the PD image from the IR image to preserve magnetization sign and applies surface coil correction (SCC) to produce the final PSIR image.
Training Objective: The network is trained end-to-end to maximize the Structural Similarity Index Measure (SSIM) between the output and the reference standard MOCO PSIR image.

C. Evaluation Protocol

Quantitative: Metrics (SSIM, PSNR, NRMSE) compared against the MOCO reference.
Qualitative: Two expert cardiologists (Level III EACVI) scored image quality on a 5-point Likert scale (1 = non-diagnostic, 5 = excellent) for 279 studies (covering all three LGE variants).
Statistical Analysis: Used exact Wilcoxon signed-rank tests for superiority and TOST (Two One-Sided Tests) for equivalence (margin = 0.25 points).

3. Key Contributions

Massive Scale & Generalizability: Unlike prior works, PSIRNet was trained on >800k slices from >55k patients across multiple sites and scanner configurations, demonstrating robust performance across Bright Blood, Dark Blood, and Wideband variants.
Physics-Guided Unrolled Reconstruction: The model integrates physical constraints (Landweber iterations, data consistency) with deep learning, allowing it to learn the reconstruction mapping without explicit IR-PD co-registration.
Single-Shot Acquisition: It successfully reconstructs diagnostic images from a single 2-heartbeat acquisition, removing the need for 8–24 averages.
Real-Time Inference: The inference time is reduced from >5 seconds/slice (CPU) to ~100 milliseconds/slice (GPU), enabling near real-time reconstruction.

4. Results

Quantitative Performance

PSIRNet achieved high similarity to the MOCO reference:
- SSIM: 92.26% ± 0.04%
- PSNR: 33.98 ± 0.03 dB
- NRMSE: 0.0474 ± 0.0002

Qualitative (Expert) Assessment

Dark Blood LGE: PSIRNet was rated statistically superior to MOCO PSIR by both readers (Conservative $P = .002$ ).
Bright Blood & Wideband LGE: PSIRNet was rated statistically equivalent to MOCO PSIR (Conservative $P < .001$ for equivalence).
Visual Quality: The network successfully suppressed noise in single-shot acquisitions, preserving the conspicuity of myocardial infarction and scar borders, even in challenging Wideband scenarios where conventional methods require up to 48 heartbeats.

Efficiency

Acquisition Time: Reduced by 8-fold (Bright Blood) to 24-fold (Dark/Wideband).
Reconstruction Time: Reduced from >5 sec/slice to 0.1 sec/slice.

5. Significance and Clinical Impact

Workflow Transformation: PSIRNet enables a shift from lengthy, multi-average scans to rapid single-shot acquisitions. This can reduce total scan time by up to 15 minutes, increasing patient throughput and reducing scanner bottlenecks.
Enhanced Imaging Capabilities: The time saved can be reinvested to:
- Acquire thinner slices for higher spatial resolution.
- Cover the entire heart volumetrically (whole-heart coverage).
- Include additional diagnostic sequences in the protocol.
Real-Time Feedback: The 100ms inference time allows technologists to verify image quality immediately, minimizing the need for patient callbacks or repeat scans due to motion artifacts.
Accessibility: By making high-quality LGE imaging faster and more robust, it improves the feasibility of cardiac MRI for patients who cannot hold their breath or tolerate long scan times.

Limitations & Future Work:
The study is retrospective, and the model was trained only on Siemens scanners. Prospective validation in a live clinical setting and testing on other vendor platforms (GE, Philips) are required before routine clinical deployment. Additionally, the study focused on image quality metrics rather than downstream clinical decision-making (e.g., scar quantification accuracy).

Availability:
The source code and trained weights are publicly available via GitHub and Hugging Face.