Skip to main content
Back to timeline
arXivSource publication:

CT-to-PET synthesis learns from its own candidates: MedicalDPO cuts BBDM's LPIPS by 11.67% with preference fine-tuning

Synopsis

Brownian Bridge diffusion models can generate several PET candidates of differing quality under the same CT condition, yet standard training aligns each sample to ground truth and leaves those relative differences unused. MedicalDPO turns rankings along spatial, intensity and structural fidelity into automatic preference pairs, then fine-tunes with a relative denoising-error preference objective jointly with the original Brownian Bridge loss. On the BBDM and CPDM backbones, LPIPS, MAE, SSIM and PSNR all improve; with multi-pair rotation on BBDM, LPIPS and MAE fall by 11.67% and 13.22%.

Source-provided article image: MedicalDPO: Multi-Objective Direct Preference Optimization for CT to PET Translation
arXiv

Interpretation

CT-to-PET models can now learn an ordering from their own generated candidates instead of only matching paired ground truth. Stochastic sampling under one CT condition yields PET images of clearly different quality, some localizing high-uptake regions well and others showing localization errors or intensity deviations, and this intra-condition quality hierarchy was previously unused by the training objective. Brownian Bridge diffusion training optimized each target sample independently and did not explicitly use the relative quality differences among candidates generated under the same condition. On BBDM, multi-pair rotation lowered LPIPS from 0.1131 to 0.0999 and MAE from 711.38 to 617.31 against the pretrained backbone, reductions of 11.67% and 13.22%; the single-pair strategy also reached LPIPS 0.1006 and MAE 626.42.

The preference pairs are not human-labeled; they are filtered automatically along three complementary fidelity dimensions: spatial fidelity via Dice on high-uptake regions, intensity fidelity via MAE and attention-weighted MAE, and structural fidelity via SSIM. Each dimension is ranked separately and combined with equal weight into an aggregate rank; only candidates in the top-5 or bottom-5 aggregate ranks and supported by at least two of the three dimensions enter the good or bad pool, and a pair is kept only when the preferred candidate beats the rejected one on all four measurements. Dice agreed with the other metrics on the best candidate only 15.9% to 21.6% of the time, whereas MAE and WMAE had a mean Spearman correlation of 0.880 and 59.8% best-candidate agreement, showing that the spatial dimension carries information the intensity and structural rankings miss; among retained pairs, preferred candidates reached Dice 0.512 versus 0.217 and MAE 0.0095 versus 0.0176.

The preference objective is trained jointly with the original Brownian Bridge reconstruction objective to anchor the pretrained CT-to-PET mapping and keep preference-only fine-tuning from pushing absolute synthesis fidelity away. In the ablation, DPO-only fine-tuning raised BBDM's LPIPS from 0.1131 to 0.1911 and MAE from 711.38 to 2143.13 while SSIM fell from 0.8405 to 0.5079; with the preservation term, BBDM under the single-pair strategy reached LPIPS 0.1006, MAE 626.42, SSIM 0.8591 and PSNR 25.02. A control that fine-tuned only on ground truth stayed close to the base models, indicating the gains are not simply from extra training; on CPDM, the single-pair strategy lowered MAE from 751.05 to 711.74, a 5.23% reduction.

Perspective

The demonstrated scope is 2D slice-level CT-to-PET synthesis, evaluated on the BBDM and CPDM Brownian Bridge backbones with four image-fidelity metrics, LPIPS, MAE, SSIM and PSNR, on a test set of 246 paired slices averaged over three random seeds.

Evaluation stays mainly at image-fidelity metrics and does not yet cover clinically relevant endpoints; synthesis is based on 2D slices rather than native 3D translation. The regularization weight was chosen by a sensitivity analysis on a 2.5K preference subset, and the feasible interval from the gradient-geometry analysis varied substantially across random subsets, so it serves as a screening criterion rather than a universal guarantee.