Leveraging transfer learning for post-operative brain tumor segmentation across MRI datasets

Motivation: Automated post-operative brain tumor segmentation is challenging due to treatment-related changes and limited availability of annotated datasets. Goal(s): This work aims to assess whether transfer learning from models pre-trained on pre-operative data would improve post-operative tumor segmentation. Approach: An ensemble segmentation model was trained on pre-operative data and then consecutively retrained on two post-operative datasets applying transfer learning to provide insights into model performance across pre- and post-operative contexts. Results: Transfer learning significantly improved post-operative segmentation accuracy, particularly on related datasets, but performance declined when testing on data introduced earlier in training, highlighting the challenges of model knowledge retention. Impact: This work addresses the hurdles of automated post-operative tumor segmentation by demonstrating that transfer learning from pre-operative models can improve post-treatment segmentation. The importance of large annotated datasets and the effects of catastrophic forgetting and model knowledge retention are highlighted.