Deep learning approach for DCE-MRI parameter estimation: Evaluating signal intensity and concentration-time...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Dynamic contrast-enhanced MRI (DCE-MRI) is crucial for brain tumor analysis using generalized tracer kinetic (GTK) modeling.
• Conventional methods like non-linear least squares (NLLS) are slow and sensitive to imaging protocol variations, affecting parameter estimation consistency.
• Existing deep learning models often overlook the impact of acquisition protocol variability.

Purpose of the Study:

• To develop and validate CNN_CON, a novel convolutional neural network for GTK parameter estimation in brain tumors.
• To assess CNN_CON's accuracy, robustness to DCE-MRI protocol variations, and computational efficiency.
• To compare CNN_CON performance against NLLS and another deep learning method (AIF-TK Net).

Main Methods:

• CNN_CON was trained on synthetic DCE-MRI data simulating protocol variations and fine-tuned on 72 glioma patient datasets.
• Validation involved 18 test patients and two external datasets (cross-scanner and cross-institutional).
• Performance metrics included mean absolute errors for kinetic parameters (K^trans, v_p, v_e) and comparison with NLLS and AIF-TK Net.

Main Results:

• CNN_CON demonstrated comparable accuracy to NLLS for GTK parameters with significantly reduced computation time (17 seconds vs. 15 minutes).
• External validation confirmed consistent performance across different scanners and institutions.
• CNN_CON showed superior accuracy (2-3x) compared to AIF-TK Net and maintained diagnostic capability for tumor grading (AUC=0.89).

Conclusions:

• The concentration-based CNN_CON approach offers robust and efficient GTK parameter estimation for DCE-MRI.
• Its validated performance across multiple centers supports consistent clinical application of perfusion imaging biomarkers.
• CNN_CON presents a promising advancement for brain tumor analysis using DCE-MRI.