Core Concepts
TransUNet model combines Transformer self-attention with U-Net for brain metastases segmentation.
Abstract
Abstract:
Brain tumors, especially metastases, pose challenges due to diverse appearances and sizes.
TransUNet model combines Transformer self-attention with U-Net for segmentation precision.
Method:
3D-TransUNet variants: Encoder-only and Decoder-only explored.
Encoder-only requires Masked-Autoencoder pre-training for better initialization.
Training Details:
MAE Pre-training accelerates convergence, improving performance.
Experiments and Results:
Encoder-only model secures second place in BraTS-METS 2023 challenge.
Conclusion:
TransUNet model shows promise for brain metastases segmentation.
Stats
Masked-Autoencoder pre-training is required for a better initialization of the Transformer Encoder and accelerates the training process.
The average lesion-wise Dice score on the test set was 59.8%.
The Decoder-only model demonstrates a substantial increase in Dice score by 2.9%.