Research Project

Overview

CycleDPM is a dual-stage diffusion model framework for anatomy-preserving CBCT-to-CT image synthesis, designed for adaptive radiation therapy. CBCT images suffer from noise and scatter artifacts that degrade their clinical utility, and CycleDPM corrects these artifacts to synthesize high-quality CT images directly from CBCT scans.

A dedicated encoder and fusion block integrate input-image features while preserving fine anatomical structures. A reduced-time sampling strategy also addresses the slow inference speed usually associated with diffusion models.

The method achieves state-of-the-art quantitative results with MAE: 67 -> 46.9 [HU] and SSIM: 0.845 -> 0.902, outperforming GAN-based baselines. The work was presented at ICNGC 2025 and is currently being prepared for broader publication.

CBCT vs Synthesized CT Visual Comparison

Compare the original CBCT slice with the synthesized CT output using a draggable divider and a direct slider control.

Synthesized CT

CBCT

Drag anywhere to compare

Comparison slider

Use the handle or the range control to compare CBCT and synthesized CT.