2892 - Exploiting Semantic Asymmetry to Enhance the Accuracy of Gross Tumor Volume Delineation in Nasopharyngeal Carcinoma on Planning CT

Purpose/Objective(s): In the radiation therapy of nasopharyngeal carcinoma (NPC), clinicians typically delineate the gross tumor volume (GTV) using non-contrast planning computed tomography to ensure accurate radiation dose delivery. However, the low contrast between tumors and adjacent normal tissues requires radiation oncologists to delineate the tumors with additional reference from MRI images manually. Here, we propose a novel approach to directly segment NPC tumors on non-contrast planning CT images, circumventing potential registration errors when aligning MRI or MRI-derived tumor masks to planning CT. To address the low contrast issues between tumors and adjacent normal structures in planning CT, we introduce a 3D Semantic Asymmetry Tumor Segmentation (SATS) method. Experiments demonstrate that the proposed SATS achieves the leading performance in internal and external testing.

Materials/Methods: We collected and curated an in-house dataset from the hospital for deep segmentation model development, which consisted of 145 NPC patients with pCT, enhanced diagnostic CT, and diagnostic MRI (T1 & T2 phases). Diagnostic CT and MRI were registered to pCT using affine and deformable transformation (DEEDS). Additionally, we collected and curated one publicly available dataset (SegRap2023) as external validation, containing 98 no-contrast pCT and enhanced CT. GTV annotations of all datasets were examined and edited by two experienced radiation oncologists following the international GTV delineation consensus guideline. Our approach first normalizes anatomical symmetry by aligning pCT scans using automatically segmented head and neck landmarks, ensuring bilateral symmetry along the central sagittal plane. This mitigates asymmetries caused by patient positioning during CT acquisition. We then introduce a Siamese contrastive learning framework, combining conventional segmentation loss with a voxel-level margin loss. The margin loss minimizes feature distances between symmetric regions in original and flipped pCT scans without tumors while maximizing distances in tumor-affected regions, enhancing sensitivity to semantic asymmetries.

Results: Experiments demonstrate that the proposed SATS achieves the leading NPC GTV segmentation performance in both internal testing, a mean Dice score of 81.2% and 95% Hausdorff distance (HD95) of 4.0mm, and external testing, with at least 2% absolute Dice score improvement when compared to state-of-the-art methods.

Conclusion: We propose a novel semantic asymmetry learning method designed to leverage the inherent asymmetrical properties of tumors in the nasopharyngeal region. The model achieves high quantitative performance, which is evaluated on the internal and external dataset.