Fully automated volumetry of ventricular subregions on computed tomography using object detection and semantic segmentation.

Background: Our goal was to develop and validate machine learning models that are capable of fully automatic identification and segmentation of frontal, temporal, and posterior horns, the body of the lateral ventricle, the third and fourth ventricle, as well as the atrium on either side. Methods: Patients shunted for hydrocephalus were included. Data from two centers was used for development/external validation, respectively. Manual labelling of ventricular subregions on computed tomography (CT) was performed. First, an object detection algorithm (YOLOv5) was trained. This allowed for precise cropping of the subregions that could then be used as input for a 2D U-Net. For comparison, a nnU-Net was also trained. Precision, recall, mean average precision 50 and 50-95 (mAP50; mAP50-95) were used as performance metrics for the YOLO algorithm. Dice score, Jaccard score, and 95th percentile Hausdorff distance assessed performance for the U-Net. Results: 80 CTs from patients at our center were included, as well as 50 from a second center. The mean age was 68.59 ± 15.89 and 75.94 ± 4.17 for the first and second centers, and 43 (52.5%) and 30 (60%) were male. MAP 50, mAP50-95 was 0.728, 0.453 for internal and 0.274, 0.124 for external validation across all classes. Best mean Dice scores were 0.92 ± 0.1 and 0.90 ± 0.05 for the body of the left lateral ventricle. Conclusions: Automatic segmentation and volumetry of ventricles including their subregions was feasible with high precision on computed tomography, potentially helping the clinical evaluation of even subtle changes in ventricular volume.