Assessing Self-supervised xLSTM-UNet Architectures for Head and Neck Tumor Segmentation in MR-Guided Applications

Abstract Radiation therapy (RT) plays a pivotal role in treating head and neck cancer (HNC), with MRI-guided approaches offering superior soft tissue contrast and daily adaptive capabilities that significantly enhance treatment precision while minimizing side effects. To optimize MRI-guided adaptive RT for HNC, we propose a novel two-stage model for Head and Neck Tumor Segmentation. In the first stage, we leverage a Self-Supervised 3D Student-Teacher Learning Framework, specifically utilizing the DINOv2 architecture, to learn effective representations from a limited unlabeled dataset. This approach effectively addresses the challenge posed by the scarcity of annotated data, enabling the model to generalize better in tumor identification and segmentation. In the second stage, we fine-tune an xLSTM-based UNet model that is specifically designed to capture both spatial and sequential features of tumor progression. This hybrid architecture improves segmentation accuracy by integrating temporal dependencies, making it particularly well-suited for MRI-guided adaptive RT planning in HNC. The model’s performance is rigorously evaluated on a diverse set of HNC cases, demonstrating significant improvements over state-of-the-art deep learning models in accurately segmenting tumor structures. Our proposed solution achieved an impressive mean aggregated Dice Coefficient of 0.81 for pre-RT segments and 0.65 for mid-RT segments, underscoring its effectiveness in automated segmentation tasks. This work advances the field of HNC imaging by providing a robust, generalizable solution for automated Head and Neck Tumor Segmentation, ultimately enhancing the quality of care for patients undergoing RT. Our team name is DeepLearnAI (CEMRG). The code for this work is available at https://github.com/RespectKnowledge/SSL-based-DINOv2_Vision-LSTM_Head-and-Neck-Tumor_Segmentation .