ESTRO 2024 - Abstract Book

S3094

Physics - Autosegmentation

ESTRO 2024

Magnetic Resonance Imaging (MRI) and Positron Emission Tomography/Computed Tomography (PET/CT) are crucial for DIL localization. PET, alongside planning CT, is consistently employed for precise DIL delineation during the treatment planning stage, offering heightened sensitivity for small or low-grade DIL detection and superior specificity compared to MRI in distinguishing DIL from benign areas. On the other hand, MR guided RT (MRgRT) has superior soft tissue contrast during image guidance or during treatment planning. MRI-guided radiation therapy systems, ie, Viewray and Unity, can image and treat cancer patients simultaneously. Daily MRI images enhance differentiation between DIL and the Gross Tumor Volume (GTV), facilitating Adaptive Radiotherapy (ART) for precise dose delivery within the target area. Nonetheless, accurate DIL delineation without daily PET images presents a major challenge. To address this challenge, we proposed an innovative boundary enhancement methodology, based on multi-modal image confidence (MMC), to achieve accurate DIL and GTV delineation in daily MR images, leveraging complementary information from pre-treatment PET and MRI for prostate SBRT.

Material/Methods:

Our algorithm consists of three main steps. Initially, deformable registration aligns DIL and GTV from pre-treatment images with the daily MRI scans. Subsequently, Trimap images are generated based on the pre-treatment delineations. Finally, the Multi-Modal Confidence (MMC) algorithm computes confidence values, ranging from 0 to 1, for each voxel, reflecting the probability of each voxel belonging to the Region of Interest (ROI). This results in two sets of confidence images for DIL and GTV, which can be readily converted into binary delineations for downstream tasks.

The MMC algorithm can be summarized as follows:

1) It constructs a Trimap by segmenting voxels into three regions: those confidently belonging to the ROI (foreground, Confidence 1), those clearly outside the ROI beyond the outer contour (background, Confidence 0), and those in an uncertain intermediate region (unknown region). An example of Trimap is visualized in Figure 1. 2) The MMC algorithm utilizes a multi-modal 3D medical image matting approach, calculating 3D spatial correlation matrices for voxels in the feature space. It estimates the confidence of uncertain voxels regarding their ROI membership based on their correlation with both foreground and background voxels from the Trimap. Importantly, this algorithm does not require model training, in contrast to deep learning (DL)-based methods, and provides deterministic results interpretable through a closed-form solution. This advantage ensures its robustness against distributional shifts between training and real-world data in deployment, distinguishing it from DL-based methods. To verify the effectiveness of the algorithm, our study included a cohort of retrospective prostate cancer patients treated with SBRT on a low filed ViewRay MRIDian system. Each patient underwent pre-treatment PET-CT scan, in addition to MRI scans, from which the DIL and GTV areas were delineated based on the PET and MR images. The DIL can be obtained through threshold segmentation or manual delineation. Furthermore, all patients underwent daily MR scans before each treatment fraction.