Organizations: Z. Li and J. Zhang are with the Shanghai Key Lab of Intelligent Information Processing, School of Computer Science and Artificial Intelligence, Fudan University, Shanghai 200433, China. · Z. Li, D. Li, D. Jin, H. Luo, and F. Wang are with Alibaba DAMO Academy, China. · C. Ma and H. Shan are with the Institute of Science and Technology for Brain-inspired Intelligence and MOE Frontiers Center for Brain Science, Fudan University, Shanghai 200433, China, and also with Shanghai Center for Brain Science and Brain-inspired Technology, Shanghai 201602, China. · Z. Feng is with the Department of Radiology, First Affiliated Hospital of College of Medical Science, Zhejiang University, Hangzhou, China.
Contrast-enhanced computed tomography (CECT) is pivotal for highlighting tissue perfusion and vascularity, yet its clinical ubiquity is impeded by the invasive nature of contrast agents and radiation risks. While virtual contrast enhancement (VCE) offers an alternative to synthesizing CECT from non-contrast CT (NCCT), existing methods struggle with anatomical heterogeneity and spatial misalignment, leading to inconsistent enhancement patterns and incorrect details. This paper introduces PHASOR, a volumetric diffusion framework for high-fidelity CT VCE. By treating CT volumes as coherent sequences, we leverage a video diffusion model to enhance structural coherence and volumetric accuracy. To ensure anatomy-phase consistent synthesis, we introduce two complementary modules. First, anatomy-routed mixture-of-experts (AR-MoE) anchors distinct enhancement patterns to anatomical semantics, with organ-specific memory to capture salient details. Second, intensity-phase aware representation alignment (IP-REPA) highlights intricate contrast signals while mitigating the impact of imperfect spatial alignment. Extensive experiments across three datasets demonstrate that PHASOR significantly outperforms state-of-the-art methods in both synthesis quality and enhancement accuracy.
Figures & tables
Fig. 1: Overview of Phasor , which takes NCCT as input and synthesizes CECT. (a) IP-REPA aligns DiT features with representations extracted from a teacher encoder during training. (b) TotalSegmentor provides organ-wise anatomy masks. (c) AR-MoE routes tokens to specialized experts based on the anatomy masks.
Dataset
Total (#)
Total
Thickness (mm)
Slice Range
Train (vol./slices)
Test (vol./slices)
VinDr-Multiphase [ 31 ]
183
168
[0.3, 5.0]
[38, 1745]
130 / 320,140
38 / 20,905
WAW-TACE [ 32 ]
233
159
[0.8, 7.5]
[36, 526]
138 / 91,225
21 / 16,890
Abdomen
900
900
[0.45, 5.0]
[42, 540]
800 / 373,647
100 / 46,519
TABLE I: Statistics of the datasets used in this study. The Abdomen denotes an in-house dataset collected from the First Affiliated Hospital, Zhejiang University School of Medicine. “Total (#)” refers to the number of raw cases, while “Total” indicates the number of valid cases after data cleaning and registration.
Fig. 2: Quantitative comparison of NCCT-to-CECT synthesis on the (a) VinDr-Multiphase and (b) WAW-TRACE datasets. “Prec.” denotes the precision score.
Methods
NC → A
NC → V
PSNR
SSIM
Dice
Recall
Prec.
FID
PSNR
SSIM
Dice
Recall
Prec.
FID
CycleGAN [ 33 ]
29.29
0.78
0.20
0.38
0.14
32.17
29.30
0.83
0.42
0.47
0.39
27.22
Pix2pix [ 34 ]
29.54
0.84
0.16
0.17
0.16
20.79
30.65
0.93
0.57
0.55
0.59
10.76
EaGAN [ 35 ]
23.47
0.59
0.29
0.35
0.26
25.97
28.89
0.82
0.38
0.44
0.34
22.88
MAFormer [ 10 ]
29.48
0.93
0.35
0.37
0.33
15.49
28.76
0.93
0.60
0.68
0.54
16.57
ALDM [ 20 ]
30.72
0.92
0.40
0.51
0.33
10.66
27.26
0.89
0.41
0.33
0.54
3.25
TABLE III: Quantitative evaluation for state-of-the-art methods on Abdomen dataset. The best results are marked in bold .
Fig. 3: Qualitative comparison of different methods on the VinDr-Multiphase and WAW-TRACE dataset. The display window is [-160, 240] HU.
Fig. 4: Qualitative comparison of different methods on the Abdomen dataset. The display window is [-160, 240] HU.
Fig. 5: 3D visualization of the subtraction volume between the synthesized CECT and the NCCT. Zoom in for better view.
Methods
NC → A
NC → V
PSNR
Dice
FID
PSNR
Dice
FID
MAFormer [ 10 ]
25.22
0.33
44.30
24.33
0.49
45.24
ALDM [ 20 ]
22.49
0.43
27.25
24.35
0.19
26.47
SMILE [ 11 ]
27.11
0.34
13.79
26.73
0.61
10.13
Phasor
29.04
0.53
6.27
29.48
0.58
4.85
TABLE IV: Cross-dataset generalization from the Abdomen dataset to the WAW-TRACE dataset without any fine-tuning.
Unregistered training data
Registered training data
3D
AR-MoE
IP-REPA
PSNR
SSIM
Dice
Recall
Prec.
FID
PSNR
SSIM
Dice
Recall
Prec.
FID
16.447
0.181
0.296
0.700
0.193
62.537
19.484
0.540
0.390
0.407
0.383
32.537
✓
23.989
0.803
0.483
0.640
0.397
10.656
27.160
0.882
0.482
0.422
0.573
7.694
✓
✓
25.203
0.826
0.492
0.520
0.483
8.364
27.665
0.895
0.539
0.591
0.501
7.735
✓
✓
24.926
0.824
0.507
0.548
0.487
8.662
27.237
0.888
0.540
0.595
0.500
7.923
✓
✓
✓
25.034
0.824
0.508
0.578
0.464
8.269
27.734
0.894
0.541
0.553
0.536
7.653
TABLE V: Ablation of components in Phasor on the VinDr-Multiphase dataset. We report results with and without deformable registration. All models are trained for 3.2k iterations. The best and second-best results are highlighted in bold and underline , respectively.
Fig. 6: Detailed ablation studies of the design of Phasor . (a) Ablation study on the design of AR-MoE, including the impact of selecting different anatomical structures for the routing mechanism, and MoE placement by inserting AR-MoE layers at different depths—shallow (layers [2,4,8]), middle (layers [14,16,18]), and deep (layers [24,26,28])—within the 30-layer DiT model. (b) Ablation on IP-REPA, using different losses and teacher models.
Fig. 7: Detailed ablation of the number of organ-specific memories L in AR-MoE. Models are trained without IP-REPA for 4k iterations.
Fig. 8: t-SNE visualization of layer-wise input and output representations of AR-MoE at different timesteps t during inference stage.
Science and Technology Organization, GE HealthCare, Bangalore, INDIA · CT Engineering, GE HealthCare, Waukesha, USA · Science and Technology Organization, GE HealthCare, San Ramon, USA +1
Center for Big Data and Intelligent Medicine, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China · Key Laboratory of Digital Health and Intelligent Medicine, Chongqing Municipal Health Commission, Chongqing, China · Chongqing Translational Medicine Center, Chongqing, China