Organs-at-Risk Segmentation on T2-Weighted Magnetic Resonance Imaging Using a Transformer-Based Model

Abstract

<h3>Purpose/Objective(s)</h3> The scarcity of high-quality, annotated MRI data for MR-only radiotherapy (RT) treatment planning remains a limitation for training automatic segmentation algorithms. In this study, we develop a framework first to synthesize T₂-weighted MRI from CT using unpaired input images, and subsequently train a volumetric transformer-based segmentation model with propagated CT contours to perform segmentation on acquired T₂-weighted MRI. <h3>Materials/Methods</h3> A retrospective database including 46 patients with locally advanced was identified. For synthesis, a 3D patch-based Cycle-GAN model (patch size of 96 × 96 × 96) was trained and validated on 13 and 5 patients respectively. Then, a transformer-based segmentation algorithm was trained and validated on synthetic MR images from 32 and 7 patients respectively to segment the femoral heads, bladder, rectum, bowel and sigmoid (binary cross entropy loss + Dice loss). <h3>Results</h3> The results from automatic contouring of OARs on synthetic images revealed that our framework delineated the femoral heads and the bladder with average Dice scores and Hausdorff distance of >0.9 and <10mm respectively. However, the segmented contours were less accurate in bowel and rectum due to their variable geometry and difficulty in segregating from adjacent soft tissues. The predictions of these organs on acquired T₂-weighted MR images also showed acceptable accuracy with visual inspection. <h3>Conclusion</h3> Our framework showed the feasibility of transferring anatomical knowledge from a pre-existing labelled CT data for RT to synthetic T₂-weighted MR images, particularly in the femoral heads and bladder. This approach provides significant clinical value as it uses an unsupervised approach for data augmentation and/or domain adaptation Future studies will quantitatively evaluate our framework's performance on larger acquired MRI cohorts and explore weakly-supervised techniques to assess performance improvements on more challenging soft tissues (e.g., bowel, rectum) with inclusion of several labelled MRI input datasets.