Abstract No. 543 Fully Automated Segmentation of Venous Malformations in Multiple Body Regions Based on a Large-Scale Magnetic Resonance Imaging Dataset

The overall purpose of this work is to quantify radiomics within Venous Malformations (VMs) and build disease progression and treatment prediction models. The quantification of radiomics necessitates the VM lesion segmentation. Therefore, the goal is to develop a fully automated algorithm to segment VMs in multiple body regions. Retrospective pre-treatment and post-treatment T2 weighted MRI scans from 200 subjects, with 100 having VMs in head and neck and 100 in extremities (hand, arm, leg, and foot) were used for this study. The ground truth VM segmentation was annotated by radiology fellows under the direct supervision of an interventional radiologist. All MRI volumes were resampled to 2 mm×2 mm×2 mm resolution and resized to 128×128×128 voxels. A 3D deep learning segmentation model, UNet++, was trained on the training set (80%) with 4-fold cross-validation and evaluated on the test set (20%). Data augmentation including volume rotation, contrast variation, and Gaussian blur was applied during training. To assess the effect of simultaneous training with multiple body regions, two experiment settings were devised: 1) two models were trained separately, one with head and neck data (head-model) and the other one with extremities data (extremity-model) and 2) a single model was trained on both head and neck data and extremities data (joint-model). The models were evaluated based on the Dice score, a metric from 0 to 1 to measure the similarity between ground truth and predicted segmentations. The mean and standard deviation of the cross-validation were reported. For the VMs larger than 8cc in the test set, the joint-model achieved a 0.66±0.03 Dice score on the head and neck (compared to 0.63±0.02 from the head-model) and 0.60±0.04 Dice score on the extremities (compared to 0.52±0.02 from the extremity-model). For the entire test set, the joint-model reached a Dice score of 0.51±0.04 on head and neck (compared to 0.48±0.02 from the head-model) and 0.47±0.03 on extremities (compared to 0.41±0.02 from the extremity-model). The universal segmentation model, trained with a large-scale dataset from multiple body regions simultaneously, surpasses the performance of models trained with those regions individually. The model's prediction tends to be more accurate for larger VMs. Future research will focus on multi-contrast training with T2, T1 pre-gad, and T1 post-gad modalities, training with missing contrasts, and signal harmonization.