Development and Practical Implementation of a Deep Learning–Based Pipeline for Automated Pre- and Postoperative Glioma Segmentation

Segmentation labels computed using the AR-CA model and 2 baseline models for comparison, shown here with the ground truth expert manual segmentation on representative preoperative (A) and postoperative (B) glioma (at 5-months’ follow-up) test cases. Segmented tumor regions overlaid on postcontrast 3D MPRAGE and 3D SPACE FLAIR. The blue mask denotes the whole tumor (peritumoral edema and tumor core); the yellow mask denotes the tumor core (nonenhancing, necrotic tumor/resection cavity, and enhancing portions of the tumor); and the red mask denotes the enhancing tumor. Arrows mark areas of overestimation or underestimation of peritumoral edema by the baseline models compared with the AR-CA model and the ground truth segmentations.

Schematic visualization of the implementation pipeline. An end-to-end pipeline was built to automate routing of relevant DICOM series from the MR imaging scanner through a vendor-neutral archive server to the inference server where preprocessing, automatic segmentation, and postprocessing tasks are executed. Thereafter, output results are sent back to the PACS for viewing using the data-transmit server again. Overall total processing time for 1 case is about 10 minutes including data routing (∼1 minute), preprocessing (∼6 minutes), segmentation (∼1–2 minutes), and postprocessing (∼1 minute).

AR-CA segmentation model performance for high-grade gliomas with (right) and without (left) skull-stripping. T1 MPRAGE postcontrast (upper) and FLAIR (lower) images are shown. This example demonstrates how preprocessing steps are necessary to facilitate proper segmentation with obvious errors in estimating the whole tumor and entirely failing to segment the tumor core and enhancing tumor subregions.