Improved quantitative parameter estimation for prostate T₂ relaxometry using convolutional neural networks.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Computational Biology

Background:

• Quantitative parameter mapping in MRI typically uses curve fitting.
• Estimating T2 parameters is crucial for prostate imaging.
• Current methods face challenges, particularly in low signal-to-noise regions.

Purpose of the Study:

• To compare conventional curve fitting techniques with neural network (NN) methods for T2 measurement in the prostate.
• To evaluate the accuracy, precision, and noise robustness of different NN architectures and training strategies against established curve fitting methods.

Main Methods:

• Generation of large, physics-based synthetic datasets simulating T2 mapping acquisitions for NN training and performance comparison.
• Implementation and comparison of four NN combinations (architectures and training corpora) against four curve fitting strategies.
• Quantitative evaluation using synthetic data with known ground truth and in vivo data with noise augmentation.

Main Results:

• A convolutional neural network (CNN) trained on naturalistic synthetic data demonstrated the highest accuracy and precision on synthetic datasets.
• The best-performing CNN produced low-noise T2 maps on in vivo data.
• This CNN method exhibited the least performance degradation with increasing input noise levels.

Conclusions:

• Supervised training of a CNN using synthetic data can yield superior T2 estimation performance compared to conventional curve fitting.
• NN-based T2 mapping shows particular promise for improving accuracy in low signal-to-noise ratio regions of the prostate.
• This approach offers a potential advancement for quantitative MRI in prostate imaging.