SPINNED: Simulation-based physics-informed neural network for deconvolution of dynamic susceptibility contrast MRI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Dynamic susceptibility contrast (DSC) MRI is crucial for assessing brain perfusion.
• Accurate deconvolution is essential for quantifying DSC-MRI parameters.
• Existing deconvolution methods face challenges with accuracy and robustness.

Purpose of the Study:

• To introduce the simulation-based physics-informed neural network for deconvolution of DSC-MRI (SPINNED).
• To establish SPINNED as a robust and accurate alternative to current deconvolution techniques.
• To enhance the reliability of DSC-MRI data analysis.

Main Methods:

• Developed SPINNED using synthetic DSC-MRI data generated from simulated tissue residue and arterial input functions.
• Trained the SPINNED model to learn the deconvolution relationship within DSC-MRI data.
• Validated SPINNED against circulant and Volterra singular value decomposition using simulated and real patient data.

Main Results:

• SPINNED demonstrated superior accuracy across all signal-to-noise ratio (SNR) levels compared to conventional methods.
• The method exhibited enhanced robustness against noise in both simulated and real-world patient data.
• SPINNED achieved significantly faster processing speeds than existing deconvolution techniques.

Conclusions:

• SPINNED presents a viable alternative for DSC-MRI deconvolution, outperforming current approaches.
• The method eliminates the need for separate ground-truth measurements, simplifying training.
• SPINNED offers faster processing and broader clinical applicability, promising more reliable and rapid diagnoses.