Deep learning approach for DCE-MRI parameter estimation: Evaluating signal intensity and concentration-time curve-based convolution-neural-networks

Abstract

Dynamic contrast-enhanced MRI (DCE-MRI) is common technique for assessing tissue perfusion and permeability in brain tumors (e.g., gliomas), using generalized tracer kinetic (GTK) analysis. However, conventional non-linear least squares (NLLS) methods are computationally expensive and sensitive to noise and protocol variations (e.g., temporal resolution, flip angle), leading to inconsistent parameter estimates. Most recent deep learning methods use signal-intensity and pre-contrast T1, without addressing acquisition protocol variation effects. We propose CNN_CON, a concentration curve-based convolutional neural network-based approach for GTK parameter estimation, designed to adapt to various DCE-MRI protocols with faster processing and investigate its accuracy and robustness to protocol variations. CNN_CON was trained on synthetic data with protocol variations, and fine-tuned on 72 glioma patients(grade 2–4). Validation was performed on 18 test patients and two external datasets: cross-scanner(n = 9, 1.5 T) and cross-institutional(n = 6, different hospital). Performance was compared against NLLS and AIF-TK Net. CNN_CON achieved mean absolute errors of 111 ± 70 × 10⁻⁵ min⁻¹, 134 ± 53 × 10⁻⁵, and 133 ± 57 × 10⁻⁴ for $K^{\mathrm{trans}}$, $v_p$ and $v_e$, respectively, with no significant difference from NLLS(p > 0.05), while reducing computational time from 15 min to 17 s. External validation demonstrated consistent performance across scanner and institutions. CNN_CON showed 2-3× better accuracy than AIF-TK Net and equivalent diagnostic capability for tumor grading (AUC = 0.89). Correlation with NLLS exceed 0.97 in normal tissues. The concentration-based approach provides superior robustness to protocol variations, validated across multiple centers, enabling consistent application of perfusion imaging biomarkers in clinical practice.