Leveraging General-purpose Models for Enhanced Head and Neck Tumor Segmentation

Identifier:  TFM:1883

Handle: https://hdl.handle.net/20.500.11797/TFM1883

Authors:  Rodriguez Llana, Sergio

Abstract:
This thesis focuses on improving the segmentation of Head and Neck tumors, a challenging task due to the region's complex anatomy and tumor variability. It evaluates the nnU-Net model, identifying its strengths and limitations. To enhance performance, pre-trained models STU-Net-B and STU-Net-H are introduced, with STU-Net-H significantly improving segmentation accuracy. The thesis proposes a novel framework, nnSAM-3D, which integrates the SAM-Med3D encoder with nnU-Net to leverage multimodal CT and PET data. This approach reaches competitve performance while keeping a low number of model trainable paramters. The work highlights the potential of pre-trained models and suggests future research for better outcomes.