The Pycerr Toolbox for the Extraction of Contrast Uptake Features from DCE-MRI
Abstract
Purpose
Dynamic Contrast-Enhanced (DCE) MRI provides functional information on tumor vascularity and perfusion, useful in characterizing tumors and gauging treatment response. We introduce an open-source extension to pyCERR for extracting non-parametric features of contrast uptake across a patient cohort and demonstrate its application in assessing breast cancer response to neoadjuvant chemotherapy (NACT).
Methods
Input images (DICOM/NIfTI) are automatically sorted into volumes, grouped by acquisition time. Longitudinal signal intensity curves are extracted from the region of interest, and a batch-processing utility facilitates the specification of the bolus arrival time (BAT) for temporal alignment across the cohort. Signal intensities are converted to contrast agent concentration or relative enhancement over baseline (mean prior to BAT). Adaptive smoothing is implemented using a Savitzky-Golay filter tuned to the estimated local noise. Resampling is performed using the Fast-Fourier Transform for uniform temporal resolution. 12 non-parametric features including the times to peak (TTP) and half-peak (TTHP), wash-in and wash-out gradients, and areas under the curve (AUC) are computed as defined by Lee et al. (IJROBP, 2018). Voxel-wise feature maps can be visualized using the integrated Napari viewer or exported to standard file formats for further analysis.
Results
Longitudinal DCE-MRI scans of 37 breast cancer patients acquired before NACT (visit V1), after the first cycle (V2), mid-treatment (V3), and post-treatment (V4) were analyzed. Scans were smoothed using a 3x3 gaussian kernel and uptake curves from expert-delineated tumors were used for feature extraction. Median absolute deviations of the initial gradient across visits V2-V4 were significantly different between pathologic complete responders (pCR) and non-pCR (p<0.05) per the Mann-Whitney U-test. A Jupyter notebook demonstrating feature extraction is distributed via Github.
Conclusion
We developed an open-source tool to extract non-parametric features of contrast uptake from DCE-MRI. Integrating features like TTHP with radiomics pipelines could enhance outcomes modeling beyond standard IBSI texture filters.