TLDR: A new MRI reconstruction method, LoSP-Prompt, addresses motion artifacts in multi-shot diffusion MRI by combining physics-informed modeling with synthetic data-tuned prompt learning. It achieves twice the spatial resolution of clinical single-shot DWI, enhances liver lesion conspicuity, and generalizes across seven diverse anatomical regions with a single model. Validated across over 10,000 clinical images, it significantly improves image quality, artifact suppression, and noise reduction, offering a robust and interpretable solution for high-resolution, multi-organ tumor diagnosis.
Magnetic Resonance Imaging (MRI), particularly Diffusion-Weighted Imaging (DWI), is a powerful tool for detecting water molecule movements in the body, which is crucial for diagnosing tumors in areas like the brain and abdomen. While traditional single-shot DWI is commonly used, multi-shot interleaved echo planar imaging (ms-iEPI) sequences offer superior resolution, better signal quality, and less geometric distortion. However, the clinical use of multi-shot DWI for body-wide tumor diagnostics has been severely limited by motion-induced phase artifacts, especially in the abdomen, caused by physiological movements like respiration, heartbeat, and intestinal peristalsis.
These complex, non-rigid movements in abdominal organs create high-order motion-induced phases that existing multi-shot DWI reconstruction methods, often designed for the more stable brain, cannot effectively handle. This leads to significant image artifacts and makes it difficult to distinguish tumor structures from noise, hindering accurate diagnosis.
Introducing LoSP-Prompt: A Novel Solution
A new reconstruction framework called LoSP-Prompt has been developed to address these challenges. This innovative method combines physics-informed modeling with a unique synthetic-data-driven prompt learning approach to deliver robust, high-resolution multi-organ DWI. The framework consists of two main parts: LoSP and Prompt-Net.
The first part, LoSP (Locally Smooth Phase), is a 1D low-rank optimization method. It cleverly breaks down the complex 2D DWI image reconstruction problem into multiple simpler 1D signal recoveries along both readout and phase encoding directions. This decomposition is key to handling the organ-specific, high-order motion phases found in the abdomen, effectively isolating and managing the rank increase caused by these complex movements.
The second part is a Prompt-Net, which uses prompt learning. This neural network, a modified ResNet18, is trained exclusively on synthetic abdominal DWI data that accurately mimics physiological motion. Its critical role is to automatically determine the optimal “saved ranks” – a parameter that dictates how much detail is retained in the 1D signal recovery. By learning from synthetic data, Prompt-Net eliminates the need for extensive real-world data supervision, making the process more robust and adaptable to diverse clinical scenarios.
Also Read:
- Advancing Brain Tumor Detection Without Manual Labels: A New Unsupervised Approach
- Enhancing Disease Risk Prediction with Patient History: A Deep Learning Approach
Remarkable Performance and Clinical Potential
LoSP-Prompt has been rigorously validated across a vast dataset of over 10,000 clinical images from 43 subjects, 4 scanner models, and 5 medical centers. The results are highly promising:
- It achieved twice the spatial resolution of standard clinical single-shot DWI, significantly enhancing the visibility of liver lesions.
- A single LoSP-Prompt model demonstrated remarkable generalizability, working effectively across 7 diverse anatomical regions, including the liver, kidney, sacroiliac, pelvis, knee, spinal cord, and brain, without any modifications or retraining.
- Radiologist evaluations (by 11 experienced radiologists) consistently showed that LoSP-Prompt outperformed state-of-the-art methods in image quality, artifact suppression, and noise reduction, often achieving “excellent” ratings.
- The method also provided better quantitative consistency for Apparent Diffusion Coefficient (ADC) values, a crucial biomarker for assessing tumor aggressiveness and treatment response.
- In patient data with liver lesions, LoSP-Prompt offered clearer lesion detectability and lower image artifacts, even surpassing navigator-based methods that require additional scans.
While LoSP-Prompt represents a significant leap forward, the researchers acknowledge two main limitations: its current ability to handle cross-slice motion and a relatively longer reconstruction time (though the Prompt-Net’s prediction itself is very fast). Despite these, the scanner-agnostic performance and robust reconstruction capabilities of LoSP-Prompt signify its transformative potential for precision oncology and body-wide tumor diagnosis. For more technical details, you can refer to the full research paper here.
This work pioneers a new framework that enables robust magnetic resonance image reconstruction through prompt and synthetic data learning, paving the way for more accurate and reliable medical imaging.
