Adaptive Plane Reformatting for 4D Flow MRI using Deep Reinforcement Learning

Javier Bisbal, Julio Sotelo, Cristóbal Arrieta, Pablo Irarrázaval, Marcelo E. Andía, Denis Parra, M.d. Valdés, Cristián Tejos, Julio García, José F. Rodríguez‐Palomares, Francesca Raimondi, Sergio UribeProceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition 2024 · DOI

TL;DR

This paper introduces AdaPR, a novel deep reinforcement learning (DRL) framework for automatically reformatting planes in 4D Flow MRI. AdaPR uses a local coordinate system to achieve robust and orientation-independent plane placement, outperforming existing DRL methods and matching the accuracy of manual measurements.

What Problem Does This Paper Address?

Analyzing 4D Flow MRI data, which provides detailed information about blood flow, requires carefully positioning specific planes within the heart and major vessels. The current standard for this "plane reformatting" is manual, which is time-consuming, subjective, and prone to inter-observer variability. Existing automated methods, including earlier DRL approaches, often struggle with variations in patient anatomy, scanner differences, and arbitrary volume orientations, limiting their real-world applicability. This paper aims to create an automated system that is accurate, robust to these variations, and as reliable as manual methods.

Methods

The core of AdaPR is a Deep Reinforcement Learning (DRL) approach, specifically using the Asynchronous Advantage Actor-Critic (A3C) algorithm. Unlike previous DRL methods that used a fixed global coordinate system, AdaPR introduces a novel local coordinate system. This allows the DRL agent to navigate and adjust the plane's position and orientation relative to the local anatomy, rather than a fixed global reference. This adaptive approach makes it resilient to variations in volume position and orientation. The agent learns by interacting with a simulated environment, receiving rewards for correctly reformatted planes and penalties for errors. The system uses 3D PC-MRA and systolic velocity data as input.

Datasets

The study utilized a dataset of 88 anonymized 4D Flow MRI scans. These scans were acquired from multiple vendors (GE, Siemens, Philips) and institutions, including both healthy volunteers and patients with congenital heart defects (such as coarctation of the aorta, tetralogy of Fallot, and bicuspid aortic valve). The data was pre-processed to an isotropic resolution for training and evaluation.

Key Insights and Ideas

  • Local Coordinates are Key: The most significant insight is that using a local, adaptive coordinate system for DRL plane reformatting dramatically improves robustness and generalization compared to global coordinate systems. This allows the agent to "understand" the local anatomy rather than relying on fixed spatial references.
  • A3C for Stability: The choice of the A3C algorithm over DQN proved crucial for stable and consistent training, especially with the added complexity of the local coordinate system.
  • Robustness to Perturbations: AdaPR demonstrates remarkable resilience to rigid transformations (rotations and translations) of the input volumes, a critical factor for real-world clinical data where scans are rarely perfectly aligned.
  • Accuracy Matches Humans: The system achieves flow quantification accuracy comparable to that of experienced human observers, bridging the gap between automated and manual methods.

Key Results

  • Superior Accuracy: AdaPR achieved a mean angular error of 6.32° ± 4.15° and a distance error of 3.40 ± 2.75 mm, outperforming baseline DRL methods like DQN-VanillaPR (11.3° ± 8.44° angular error, 6.29 ± 6.01 mm distance error).
  • Orientation Independence: AdaPR maintained consistent accuracy across various volume orientations and positions, with errors deviating only slightly from the unperturbed baseline.
  • Comparable to Manual: Flow measurements from AdaPR planes showed no significant differences compared to two manual observers, with excellent correlation (R² = 0.972 vs. Observer 1, R² = 0.968 vs. Observer 2), comparable to inter-observer agreement (R² = 0.969).
  • Efficiency: Pre-processing for a single scan took 3-5 seconds, and evaluation per plane was 3-4 seconds.

Limitations and Future Work

  • Training Data Dependency: The framework relies on supervised training, meaning accuracy is dependent on the size and diversity of the training data and the quality of manual labels. The current dataset, while diverse, had a smaller proportion of patient scans, leading to slightly larger errors in patients.
  • Pre-processing Requirements: Steps like contrast enhancement and interpolation to isotropic resolution are still needed for new datasets, which might discard some information.
  • Future Directions: The authors suggest incorporating additional training objectives (e.g., maximizing through-plane flow) to reduce biases from limited datasets. Collaborative agent approaches could further mitigate variability across planes. They also propose expanding evaluations to more complex cardiac anatomies and exploring applications beyond 4D Flow MRI, such as other MRI modalities or CT angiography.

Summary updated

Abstract

Motivation: The standard approach for plane reformatting in 4D flow MRI is manual, leading to time-consuming and user-dependent results. Goal(s): Our goal was to enhance plane reformatting in 4D flow MRI and overcome limitations associated with existing automated methods. Approach: We introduce a novel approach that employs deep reinforcement learning (DRL) with a flexible coordinate system for precise and adaptable plane reformatting. Results: Results demonstrate superior performance compared to baseline DRL and similar outcomes compared to those of landmark-based techniques, showing its potential for use in complex medical imaging scenarios beyond 4D flow MRI. Impact: The proposed framework allows for automated, precise, and adaptive plane reformatting, facilitating the use of 4D flow MRI in clinical routines. It was trained with data sets from different vendors, making this approach widely applicable.
BibTeX
@inproceedings{bisbal2024adaptive,
  author = {Javier Bisbal and Julio Sotelo and Cristóbal Arrieta and Pablo Irarrázaval and Marcelo E. Andía and Denis Parra and M.D. Valdés and Cristián Tejos and Julio García and José F. Rodríguez‐Palomares and Francesca Raimondi and Sergio Uribe},
  title = {{Adaptive Plane Reformatting for 4D Flow MRI using Deep Reinforcement Learning}},
  booktitle = {Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition},
  year = {2024},
  doi = {10.58530/2024/2232},
}
Figure