[Paper Review] Synthetic Perfusion Maps: Imaging Perfusion Deficits in DSC-MRI with Deep Learning

Andreas Hess, Raphael Meier+6 more|arXiv (Cornell University)|Jun 11, 2018

Advanced MRI Techniques and Applications3 references4 citations

TL;DR

This paper proposes a deep learning-based, model-free method for generating synthetic perfusion maps in dynamic susceptibility contrast-enhanced (DSC-MRI) using end-to-end convolutional neural networks trained solely on raw perfusion data. The approach achieves clinically comparable results to standard perfusion maps, with reduced noise and faster inference, while incorporating a bolus characterization module to handle variable contrast arrival times without requiring arterial input function (AIF) estimation.

ABSTRACT

In this work, we present a novel convolutional neural net- work based method for perfusion map generation in dynamic suscepti- bility contrast-enhanced perfusion imaging. The proposed architecture is trained end-to-end and solely relies on raw perfusion data for inference. We used a dataset of 151 acute ischemic stroke cases for evaluation. Our method generates perfusion maps that are comparable to the target maps used for clinical routine, while being model-free, fast, and less noisy.

Motivation & Objective

• To develop a model-free, end-to-end deep learning framework for generating perfusion maps directly from raw DSC-MRI data.
• To eliminate the need for manual or automated arterial input function (AIF) estimation, which is error-prone and time-consuming.
• To improve the robustness of perfusion map generation in acute ischemic stroke by handling variable contrast bolus arrival times.
• To produce perfusion maps that are less noisy and faster to compute than traditional methods, enabling real-time clinical use.

Proposed method

• A 3D U-Net-like convolutional neural network architecture is used to predict T_max perfusion maps from raw DSC-MRI sequences.
• The model processes each voxel's time series independently, relying on local temporal patterns in the perfusion signal.
• A dedicated bolus characterization structure (BCS) extracts features from a fixed anatomical region (basilar to posterior cerebral artery) to estimate bolus arrival time and AIF characteristics.
• Temporal data augmentation is applied by randomly shifting the perfusion sequence by -5 to 30 frames, with reflection-based padding to maintain signal integrity.
• The loss function uses weighted mean absolute error (MAE) with higher weights on values that most influence the MAEC metric, improving prediction accuracy on critical regions.
• Spatial and temporal normalization, including voxel-wise Gaussian smoothing and standardization, are applied during pre-processing to stabilize training.

Experimental results

Research questions

• RQ1 Can a deep learning model generate clinically relevant perfusion maps directly from raw DSC-MRI data without requiring AIF estimation or parametric modeling?
• RQ2 How does the model perform in comparison to standard clinical perfusion maps in terms of accuracy, noise, and robustness to variable bolus arrival times?
• RQ3 To what extent does temporal data augmentation improve generalization across cases with differing contrast arrival delays?
• RQ4 Does the inclusion of a dedicated bolus characterization module enhance model performance compared to a standard CNN without such inductive bias?
• RQ5 Can the model estimate prediction uncertainty, and does this uncertainty correlate with actual prediction error?

Key findings

• The proposed model achieved comparable performance to standard clinical perfusion maps, with significantly reduced noise and smoother appearance in predicted T_max maps.
• The model’s performance was robust to variable contrast bolus arrival times, as demonstrated by improved generalization through temporal data augmentation.
• Removal of the bolus characterization structure or loss weighting led to a significant drop in prediction accuracy, confirming their critical role in model performance.
• The model’s predicted uncertainty (σ²) correlated with predicted value magnitude rather than error, suggesting difficulty in predicting high-value voxels accurately due to noise in target maps.
• In cases with weak signal attenuation and high noise (e.g., due to head motion), the model failed, indicating limitations when the bolus characterization structure could not reliably detect the bolus.
• Model B, which used data augmentation during training, outperformed model A on the test set, indicating that augmentation improved generalization on unseen data with variable bolus timing.