Uncertain Feature-refinement Attention Unet: Considering Suitable Convolutional Neural Network Model for Real-time...

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Biology

Background:

• Markerless tumor tracking (MTT) using deep learning models faces challenges due to domain shifts caused by noise and anatomical variations.
• Accurate tumor segmentation is crucial for effective radiotherapy and treatment planning.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) model for real-time MTT segmentation.
• To address domain shifts in deep learning models for improved MTT accuracy.

Main Methods:

• Proposed the Uncertain Feature-refinement Attention Unet (UFA-Unet), designed to handle domain shifts between digitally reconstructed radiographs (DRRs) and kV X-ray fluoroscopic (XF) images.
• Conducted qualitative ablation studies, quantitative evaluations on lung cancer cases, and phantom studies to assess model performance and robustness.
• Compared UFA-Unet against established models like U-Net, Attention-Unet, and Swin-Unet.

Main Results:

• Ablation studies confirmed that UFA-Unet components effectively suppress over-activation, enhancing segmentation accuracy.
• Quantitative studies showed UFA-Unet's superior performance over conventional models on noisy DRRs from different treatment plans.
• Phantom studies demonstrated UFA-Unet's robust tracking capabilities across unseen respiratory phases, with a 95th percentile 3D error of 0.61-3.13 mm.

Conclusions:

• UFA-Unet achieves accurate, robust, and real-time segmentation for markerless tumor tracking.
• The model's ability to overcome domain shifts makes it suitable for clinical MTT applications.