Cone-Beam CT Segmentation for Intraoperative Electron Radiotherapy Based on U-Net Variants with Transformer and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• AI applications are growing in radiotherapy, but limited in intraoperative electron radiotherapy (IOERT).
• No AI solution exists for contouring cone beam CT (CBCT) images from mobile CBCT devices used in IOERT.
• Standard U-Net struggles with global context in medical image segmentation.

Purpose of the Study:

• To enhance CBCT segmentation accuracy for IOERT using advanced AI architectures.
• To evaluate the performance of hybrid U-Net models incorporating Transformer and xLSTM features against standard U-Net and manual segmentation.
• To address time-consuming treatment planning in 3D image-based IOERT through automatic contouring.

Main Methods:

• Trained and evaluated three AI architectures: standard U-Net, U-Net with Transformer features, and U-Net with xLSTM.
• Utilized 55 CBCT scans from breast cancer patients undergoing IOERT.
• Assessed segmentation performance using the Dice Similarity Coefficient (DSC).

Main Results:

• The U-Net with Transformer features achieved the highest average DSC of 0.88.
• This hybrid model showed a 5% average improvement over the standard U-Net (DSC 0.83).
• The U-Net with xLSTM demonstrated inferior performance (DSC 0.66) compared to the standard U-Net.

Conclusions:

• Hybrid U-Net architecture with Transformer features significantly improves CBCT segmentation accuracy for IOERT.
• This advancement can facilitate automatic contouring, enabling synthetic CT image generation.
• The findings help overcome IOERT's time-intensive treatment planning challenges.