2660 - A Temporal Attention-Guided 3D UNet for Multi-Center CBCT to Planning CT Synthesis in Image-Guided Radiation Therapy

Purpose/Objective(s):

Daily CBCT supports patient setup in IGRT, but limited HU fidelity and image quality hinder dose recalculation, often prompting additional planning CT (pCT) and extra radiation. Most CBCT-to-CT synthesis models process single CBCT volumes and underuse temporal cues in sequential CBCT, compromising soft-tissue and fine-structure preservation. We propose a temporal-attention 3D U-Net that fuses a 5-frame CBCT sequence to synthesize pCT, and we assess the contribution of temporal attention via ablation.

Materials/Methods:

Paired 5-frame CBCT sequences and pCT were retrospectively collected from three centers. Volumes were resampled to 128×128×128; HU were clipped to [-1000, 1000] and normalized per channel using the 0.5–99.5th percentiles. Data were split by center into train/val/test = 70/15/15%. The backbone is a 3D U-Net; a temporal-attention block computes frame weights (global average pooling + softmax) to fuse features across frames. Models: (i) TA-UNet (proposed), (ii) TA-UNet w/o temporal attention, (iii) plain 3D U-Net. Training: 150 epochs; loss = 0.7·L1 + 0.2·(1-SSIM) + 0.1·feature-variance; Adam optimizer with mixed precision. Metrics: PSNR, SSIM, HU_MAE, HU_RMSE, and Dice for bone (>200 HU), soft tissue (-100–200 HU), and air (=-100 HU).

Results:

On the multi-center test set, TA-UNet achieved PSNR/SSIM = 21.55 dB/0.7125 with HU_MAE/HU_RMSE = 92.74/191.03 HU and Dice (bone/soft/air) = 0.7947/0.4821/0.9445. Removing temporal attention reduced PSNR by 1.16 dB, increased HU_MAE by 19.88 HU, and decreased soft-tissue Dice from 0.4821 to 0.3749, indicating temporal attention is the primary driver. Versus plain 3D U-Net, TA-UNet improved SSIM (0.7125 vs 0.5415) and HU_MAE (92.74 vs 106.35 HU).

Conclusion:

The proposed temporal attention-guided 3D UNet model achieves state-of-the-art HU accuracy and anatomical structure preservation for multi-center CBCT-to-CT synthesis. The temporal attention module is the critical functional component, enabling effective fusion of sequential CBCT frame information to improve soft tissue delineation and overall synthesis quality. This model enables accurate dose calculation directly from daily CBCT scans in the IGRT workflow, with the potential to eliminate redundant planning CT scans, reduce patient radiation exposure, and facilitate widespread clinical adoption of online adaptive radiation therapy.

Table1. Full Quantitative Results of Proposed Model and Ablation Study

Table Notes: Bold values indicate the best overall clinical performance; all metrics are reported as mean values on the held-out multi-center test set.

Model Variant	PSNR_3D (dB)	SSIM_3D	HU_MAE (HU)	HU_RMSE (HU)	BONE_DICE	SOFT_DICE	AIR_DICE
Proposed Model (with Temporal Attention)	21.55	0.7125	92.74	191.03	0.7947	0.4821	0.9445
Ablation 1: Proposed Model without Temporal Attention	20.39	0.6588	111.18	218.47	0.7701	0.3749	0.9299
Ablation 2: Baseline Model (No Attention Mechanisms)	21.03	0.5415	106.35	198.50	0.7600	0.4550	0.9459