Semiautomated Workflow for Clinically Streamlined Glioma Parametric Response Mapping

Abstract

Management of glioblastoma multiforme remains a challenging problem despite recent advances in targeted therapies. Timely assessment of therapeutic agents is hindered by the lack of standard quantitative imaging protocols for determining targeted response. Clinical response assessment for brain tumors is determined by volumetric changes assessed at 10 weeks post-treatment initiation. Further, current clinical criteria fail to use advanced quantitative imaging approaches, such as diffusion and perfusion magnetic resonance imaging. Development of the parametric response mapping (PRM) applied to diffusion-weighted magnetic resonance imaging has provided a sensitive and early biomarker of successful cytotoxic therapy in brain tumors while maintaining a spatial context within the tumor. Although PRM provides an earlier readout than volumetry and sometimes greater sensitivity compared with traditional whole-tumor diffusion statistics, it is not routinely used for patient management; an automated and standardized software for performing the analysis and for the generation of a clinical report document is required for this. We present a semiautomated and seamless workflow for image coregistration, segmentation, and PRM classification of glioblastoma multiforme diffusion-weighted magnetic resonance imaging scans. The software solution can be integrated using local hardware or performed remotely in the cloud while providing connectivity to existing picture archive and communication systems. This is an important step toward implementing PRM analysis of solid tumors in routine clinical practice.

Keywords: diffusion; glioma; parametric response mapping.

Figure 1.

Software workflow is presented for processing and analysis of serial glioma diffusion data for parametric response mapping of apparent diffusion coefficients maps (PRM_ADC). Images are first processed to generate an apparent diffusion coefficient (ADC) map and tumor segmentation. Then, follow-up images are spatially aligned with baseline images. Once ADC maps are aligned, they are further processed into a parametric response map, consisting of 3-voxel classifications: increased, decreased and unchanged ADC. Summary statistics for clinical evaluation consist of the tumor-relative volume of each classification as well as mean ADC at each time point.

Figure 2.

Example final report of parametric response mapping (PRM) results from the online software platform. This provides PRM-relative volumes as summary statistics along with key PRM overlays and scatterplot for insight into distribution patterns and spatial context. In addition, measures of mean tumor ADC and tumor volume are provided.

Figure 3.

An example of automated segmentation reveals good agreement with the original manual segmentation. Manual segmentations drawn by an experienced radiologist are shown in the top row and semiautomated segmentation results in the second row. Generation of the semiautomated VOI began with the single-slice seed ROI shown in the bottom row (slice 13).

Figure 4.

Dice coefficients comparing semiautomated tumor segmentations to radiologist-defined segmentations. Maximum Dice per patient resulted in good agreement between segmentations, left; however, use of all slice ROIs individually as seed regions resulted in poor agreement, right. This indicates that seed ROI positioning is critical for an acceptable segmentation. The seed ROI should be placed in the middle of a tumor, where the tumor area is the largest.

Figure 5.

Robustness of semiautomated segmentation was evaluated using Dice coefficients between the best-slice seed segmentation and shifted-seed segmentations (A), showing little variation between segmentations. Resulting volumes were also found to remain consistent with varying segmentation seed (B).

Figure 6.

Receiver operator curve (ROC) analysis was used to compare final PRM_ADC+ results against the same subjects' previous results (n = 49). Semiautomated segmentation was performed on previously coregistered images, followed by automated PRM classification. The ROC statistics are compared between the 2 in the table.

Figure 7.

The ROC analysis was used to compare final PRM_ADC+ results against the same subjects' previous results (n = 40). The full workflow was applied to available original imaging data, consisting of semiautomated segmentation, image coregistration, and PRM classification. ROC statistics are compared between the 2 in the table.