3750 - Texture-Aware Synthesis of DIBH-CT from DIBH-CBCT for Adaptive Radiotherapy Replanning

Purpose/Objective(s): Deep Inspiration Breath-Hold (DIBH) radiotherapy for left-sided breast cancer is effective in reducing cardiac and pulmonary radiation exposure while improving target stability. However, conventional single-session DIBH-CT imaging fails to adapt to anatomical changes over the course of treatment, limiting timely plan adjustments. Additionally, current DIBH-CBCT images suffer from low quality and artifacts, resulting in suboptimal registration and generally poor performance in adaptiave radiotherapy. To address these issues, we propose a novel method to synthesize DIBH-CT images from fractionated DIBH-CBCT scans, enabling adaptive replanning.

Materials/Methods: This retrospective study included 103 left-sided breast cancer patients who underwent breast-conserving surgery followed by intensity-modulated radiotherapy (IMRT). DIBH-CBCT images from the first five treatment fractions and the planning DIBH-CT (DIBH-pCT) images were collected. The DIBH-CBCT images were acquired at end-inspiration breath-hold using Catalys optical surface-guidance. Patients were randomly divided into training (82 patients, 11,823 slices) and testing sets (21 patients, 3,029 slices) in an 8:2 ratio. To address the smoothing issue caused by L1 loss in traditional UNet/GAN models, we proposed a texture-aware generative adversarial network (GAN). The generator, designed in a U-shape, integrates CNN and Transformer modules with a wavelet-based channel-spatial attention mechanism (CBAM) to enhance high-frequency texture features.The discriminator employs a PatchGAN structure to improve local detail discrimination. Furthermore, due to differences in HU value histograms between DIBH-CBCT and DIBH-CT, histogram matching loss and radiomics feature loss were introduced during training to optimize HU value distribution consistency. The quality of the synthesized DIBH-CT (DIBH-sCT) images was evaluated using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE).

Results: Ablation studies showed progressive improvements in image quality: the baseline model (without CBAM or hybrid losses) achieved SSIM of 0.897, PSNR of 29.154, and MAE of 38.614± 22.280 HU. Adding CBAM increased SSIM to 0.932, PSNR to 32.826, MAE to 34.184 ± 22.517 HU. The proposed model, incorporating both CBAM and hybrid losses, further improved SSIM to 0.946, PSNR to 34.180, and MAE to 26.624 ± 20.307 HU (all p<0.001). The hybrid loss framework resulted in a 46.3% improvement in texture preservation and HU accuracy compared to the baseline model.

Conclusion: The proposed method successfully synthesizes high-quality DIBH-CT images from DIBH-CBCT, with significant improvements in texture detail preservation. This approach provides a reliable foundation for real-time dose re-optimization in DIBH radiotherapy and show considerable clinical potential for adaptive radiation therapy.