Exploring the Adaptability of Attention U-Net for Post-operative Brain Tumor Segmentation in MRI Scans

This study explores the adaptability of a segmentation model, originally trained on pre-operative MRI data, in post-operative recurrent brain tumor segmentation. We utilized the Attention U-Net model for this study. In pre-operative training, the model achieved a Dice Coefficient of 0.92 and an IOU of 0.86 for brain tumor MRI segmentation. Due to the surgical artifacts in post-operative data, performance reduced with Dice Coefficient of 0.54 and an IOU of 0. To improve the performance, the model's architecture is fine-tuned by introducing dilated convolutions and residual connections. This refinement yielded improvements in results, with a Dice Coefficient of 0.68 and an IOU of 0.62 in the post-operative context. This improvement underscores the need for further research to select and adapt efficient models, retrain specific layers with an extensive collection of post-operative images, and fine-tune model parameters to enhance feature extraction during the encoding phase.