Improving Glioma Segmentation in Low-Resolution Domains With Transfer Learning
Juampablo Heras Rivera, University of Washington
Training accurate tumor segmentation models only using data from the BraTS Sub-Saharan Africa (SSA) Glioma dataset is difficult due to the low quantity and resolution of the images. However, it is possible to improve model performance through the use of transfer learning methods which leverage insights gained from larger datasets, such as the BraTS23 Adult Glioma dataset. Here, we evaluate the performance of various transfer learning approaches on the task of improving tumor segmentation Dice and Hausdorff Distance (95%) scores on the BraTS SSA dataset. The transfer learning approaches assessed here include: Domain Adversarial Neural Networks, Fine Tuning (with and without freezing layer weights), and training with a combined dataset of low- and high-resolution images.