Improving the reliability of pharmacokinetic parameters at dynamic contrast-enhanced MRI in astrocytomas: A deep learning approach

Abstract

© RSNA, 2020.Background: Pharmacokinetic (PK) parameters obtained from dynamic contrast agent–enhanced (DCE) MRI evaluates the microcirculation permeability of astrocytomas, but the unreliability from arterial input function (AIF) remains a challenge. Purpose: To develop a deep learning model that improves the reliability of AIF for DCE MRI and to validate the reliability and diagnostic performance of PK parameters by using improved AIF in grading astrocytomas. Materials and Methods: This retrospective study included 386 patients (mean age, 52 years 6 16 [standard deviation]; 226 men) with astrocytomas diagnosed with histopathologic analysis who underwent dynamic susceptibility contrast (DSC)-enhanced and DCE MRI preoperatively from April 2010 to January 2018. The AIF was obtained from each sequence: AIF obtained from DSC-enhanced MRI (AIFDSC) and AIF measured at DCE MRI (AIFDCE). The model was trained to translate AIFDCE into AIFDSC, and after training, outputted neural-network-generated AIF (AIFgenerated DSC) with input AIFDCE. By using the three different AIFs, volume transfer constant (Ktrans), fractional volume of extravascular extracellular space (Ve), and vascular plasma space (Vp) were averaged from the tumor areas in the DCE MRI. To validate the model, intraclass correlation coefficients and areas under the receiver operating characteristic curve (AUCs) of the PK parameters in grading astrocytomas were compared by using different AIFs. Results: The AIF-generated, DSC-derived PK parameters showed higher AUCs in grading astrocytomas than those derived from AIFDCE (mean Ktrans, 0.88 [95% confidence interval {CI}: 0.81, 0.93] vs 0.72 [95% CI: 0.63, 0.79], P = .04; mean Ve, 0.87 [95% CI: 0.79, 0.92] vs 0.70 [95% CI: 0.61, 0.77], P = .049, respectively). Ktrans and Ve showed higher intraclass correlation coefficients for AIFgenerated DSC than for AIFDCE (0.91 vs 0.38, P , .001; and 0.86 vs 0.60, P , .001, respectively). In AIF analysis, baseline signal intensity (SI), maximal SI, and wash-in slope showed higher intraclass correlation coefficients with AIFgenerated DSC than AIFDCE (0.77 vs 0.29, P , .001; 0.68 vs 0.42, P = .003; and 0.66 vs 0.45, P = .01, respectively. Conclusion: A deep learning algorithm improved both reliability and diagnostic performance of MRI pharmacokinetic parameters for differentiating astrocytoma grades