A Multicenter Study on Deep Learning for Glioblastoma Auto-Segmentation with Prior Knowledge in Multimodal Imaging

Abstract

A precise radiotherapy plan is required to ensure accurate delineation of gross tumor volumes (GTV) and clinical target volumes (CTV1 and CTV2) of glioblastomas (GBMs). However, traditional manual delineation is labor intensive and highly dependent on oncologists' experience. To construct and evaluate a deep learning-based automatic delineation method using prior knowledge in multimodal medical imaging to automate precise GTV, CTV1 and CTV2 contouring in GBM patients. We retrospectively collected the CT and MRI scans of 55 eligible patients with histologically proven high-grade glioma (HGG) from an institute, these scans were performed with non-enhanced CT (CT), contrast-enhanced T1-weighted (T1C) and T2-FLAIR (T2F) sequences. We proposed a two-stage automatic segmentation framework (PKMI-Net) for GTV, CTV1 and CTV2 based on deep learning using prior knowledge in multimodal medical imaging, and its segmentation performance was evaluated with dice similarity coefficient (DSC), 95% Harsdorff distance (HD95), average surface distance (ASD) and relative volume difference (RVD). To further investigate the generalizability of our method, we designed and conducted two evaluation strategies (Mix and Cross) on four multicenter datasets (including 55 patients, 37 patients, 21 patients and 35 patients). The evaluation results with an 11-patient test set from the single institute were summarized in Table 1, the proposed method demonstrated the best accuracy in segmenting, respectively, GTV, CTV1 and CTV, achieving a DSC of 0.94, 0.95 and 0.92; HD95 of 2.07 mm, 1.18 mm and 3.80 mm; ASD of 0.69 mm, 0.39 mm and 1.13 mm and RVE of 5.50%, 3.97% and 9.68%. In the multicenter evaluation, the segmentation performance of our method implemented with the Cross strategy was comparable to that with the Mix strategy, demonstrating that our method had high and stable generalizability across multicenter datasets in automatically segmenting GTV, CTV1 and CTV2. Our proposed method achieved promising results in automatically segmenting gliomas across various datasets, which could improve the quality and efficiency of glioblastoma radiotherapy.