Abstract Radiation therapy is one of the most frequently applied cancer treatments worldwide, especially in the context of head and neck cancer. Today, MRI-guided radiation therapy planning is becoming increasingly popular due to good soft tissue contrast, lack of radiation dose delivered to the patient, and the capability of performing functional imaging. However, MRI-guided radiation therapy requires segmenting of the cancer both before and during radiation therapy. So far, the segmentation was often performed manually by experienced radiologists, however, recent advances in deep learning-based segmentation suggest that it may be possible to perform the segmentation automatically. Nevertheless, the task is arguably more difficult when using MRI compared to e.g. PET-CT because even manual segmentation of head and neck cancer in MRI volumes is challenging and time-consuming. The importance of the problem motivated the researchers to organize the HNTSMRG challenge with the aim of developing the most accurate segmentation methods, both before and during MRI-guided radiation therapy. In this work, we benchmark several different state-of-the-art segmentation architectures to verify whether the recent advances in deep encoder-decoder architectures are impactful for low data regimes and low-contrast tasks like segmenting head and neck cancer in magnetic resonance images. We show that for such cases the traditional residual UNet-based method outperforms (DSC = 0.775/0.701) recent advances such as UNETR (DSC = 0.617/0.657), SwinUNETR (DSC = 0.757/0.700), or SegMamba (DSC = 0.708/0.683). The proposed method (lWM team) achieved a mean aggregated Dice score on the closed test set at the level of 0.771 and 0.707 for the pre- and mid-therapy segmentation tasks, scoring 14th and 6th place, respectively. The results suggest that proper data preparation, objective function, and preprocessing are more influential for the segmentation of head and neck cancer than deep network architecture.