Source-linked AI summary
Diffusion-Based Synthesis of Complete 3D Biparametric Prostate MRI
Claudia Giardina, Montse Pardàs, Verónica Vilaplana
TL;DR
Limited availability of large, representative prostate MRI datasets motivates synthetic bpMRI. The paper develops anatomy-conditioned 3D diffusion models for sequential and joint synthesis. Both pathways were feasible, while direct ADC/cHBV synthesis outperformed deriving those modalities from synthetic diffusion-weighted images.
Problem
Deep-learning performance and generalization depend on large, representative annotated datasets, motivating interest in synthetic prostate MRI for dataset augmentation.
Source: §1 · pp. 1–2Method
A conditional 3D diffusion framework generates complete prostate bpMRI from anatomical masks, either sequentially with separate models or jointly in one model.
Source: §1 · p. 2, §2.1 · p. 3Results
Joint synthesis generated complete bpMRI with modality-specific distributional differences, while direct ADC/cHBV synthesis outperformed derivation from synthetic LBV/HBV.
Source: §4 · p. 9Takeaways & Limitations
For diffusion-derived targets, the reported distributional results favor direct ADC/cHBV synthesis over deriving them from synthetic LBV/HBV.
Source: §4 · p. 9Whether generated images preserve diagnostically relevant characteristics or provide clinical value remains to be validated.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Synthetic biparametric MRI (bpMRI) may help alleviate the limited availability of large and representative prostate MRI datasets for deep learning applications. In this work, we present a conditional 3D diffusion framework for anatomy-conditioned synthesis of complete prostate bpMRI from anatomical masks. The framework supports sequential synthesis, in which T2w and diffusion modalities are generated by separately trained models, and joint synthesis, in which all bpMRI modalities are generated simultaneously within a single model. Five configurations were trained using PI-CAI, FastMRI Prostate, and PROSTATEx datasets to investigate different conditioning schemes, target modalities, and synthesis strategies. Evaluation using distribution-based metrics, paired-image similarity, and qualitative assessment demonstrated the feasibility of both synthesis pathways. Joint synthesis achieved competitive distributional fidelity while generating complete bpMRI, whereas direct synthesis of ADC and computed high-b-value DWI outperformed their derivation from synthetic diffusion-weighted images.
1 Introduction
The paper presents anatomy-conditioned synthesis of complete prostate bpMRI to address limited availability of large, representative datasets for deep learning. It investigates both sequential and joint generation, including alternative ways to synthesize diffusion information.
1 Introduction
- Deep-learning model performance and generalization depend on large, representative annotated datasets, motivating prostate MRI synthesis to augment datasets with clinically relevant synthetic images.The cited applications include lesion detection, segmentation, and clinically significant prostate cancer classification.Source: §1 · pp. 1–2
- Prior work demonstrated prostate MRI synthesis with 2D diffusion models and 3D methods for individual modalities, clinically conditioned T2w, and multimodal generation.The cited 3D work includes anatomy-conditioned T2w and generation of T2w, high-b-value DWI, and ADC.Source: §1 · p. 2
- Anatomy-conditioned generation of complete 3D bpMRI from prostate-zone and lesion masks remains largely unexplored, as does the choice between acquired-image and biomarker-level diffusion synthesis.The alternatives are to generate diffusion-weighted images and derive ADC and computed high-b-value DWI, or to synthesize ADC and computed high-b-value DWI directly.Source: §1 · p. 2
- The framework generates complete prostate bpMRI from anatomical masks and evaluates sequential and joint synthesis across different diffusion targets.Five configurations were trained using PI-CAI, FastMRI Prostate, and PROSTATEx data.Source: §1 · p. 2
2 Methodology
The methodology adapts a conditional 3D diffusion model for anatomy-conditioned synthesis, evaluates five configurations across three datasets, and assesses distributional and paired-image similarity.
2.1 Framework Overview
- The conditional 3D diffusion framework predicts the clean target volume and regulates through-plane resolution while supporting flexible conditioning and single- or multi-modality targets.Anatomical masks can be used alone or with T2w; separately trained models support sequential synthesis, while a single model can generate all target modalities jointly.Source: §2.1 · pp. 2–3
2.2 Datasets and Image Preprocessing
- PI-CAI, FastMRI Prostate, and PROSTATEx supplied the MRI volumes and anatomical annotations, which were registered as needed, resampled, cropped, and intensity-normalized.The usable datasets comprised 1,158 PI-CAI, 235 FastMRI, and 281 PROSTATEx studies.Source: §2.2 · p. 3
2.3 Experimental Design
- Five configurations compared sequential and joint synthesis of acquired LBV/HBV images versus direct ADC/cHBV targets.Sequential synthesis uses a mask-conditioned T2w model followed by a model conditioned on the mask and generated T2w; joint synthesis generates T2w and diffusion targets together.Source: §2.3 · p. 4, Table 1 · p. 4
2.4 Training Details
- Models used Adam and direct x0 prediction with a voxel-wise L1 loss, combining modality-specific losses with empirical weights for multi-target models.Mask smoothing was applied stochastically during training to soften anatomical-condition boundaries.Source: §2.4 · pp. 4–5
2.5 Evaluation
- Distributional fidelity was assessed with FID, KID, Precision, and Recall; paired PSNR and SSIM were computed within the prostate region against matched real targets.For mask-only models, paired metrics indicate case-specific anatomical consistency rather than reconstruction fidelity.Source: Distribution-based evaluation · p. 5, Paired-image evaluation · p. 5
3 Results
Both joint and sequential synthesis produced competitive results. Direct ADC/cHBV synthesis outperformed deriving those modalities from synthetic LBV/HBV, while paired-image comparisons sometimes favored sequential models.
T2w
- M1 had the lowest T2w FID and KID, while joint M4/M5 achieved higher Precision and Recall and comparable distribution-based performance.Joint models also preserved comparable case-specific agreement to the dedicated T2w model.Source: §3 · pp. 5–6, Table 2 · p. 6
- M1 generally produced sharper T2w boundaries and more detailed texture than smoother, mask-consistent joint outputs from M4 and M5.Source: §3 · p. 5, Figure 2 · p. 6
LBV and HBV
- M2 had higher LBV SSIM (0.72 vs. 0.63) and HBV PSNR/SSIM (17.11/0.57 vs. 15.74/0.53), while M4 slightly led LBV PSNR (20.41 vs. 20.14).Overall, joint synthesis better matched the real image distribution, whereas sequential synthesis preserved stronger case-specific correspondence.Source: §3 · p. 6
ADC and cHBV
- Directly synthesized ADC/cHBV from M3 and M5 had better distribution-based performance than M4-derived images across metrics except cHBV Precision.The M4-derived images were computed from synthetic LBV/HBV; the authors suggest errors in those acquisitions may propagate to derived images.Source: Table 3 · p. 7, §3 · p. 7
LBV and HBV
- M4 outperformed sequential M2 on LBV/HBV distribution metrics with lower FID and KID and higher Precision; Recall was higher for LBV but slightly lower for HBV.The paired-image comparison favored M2 on LBV SSIM and HBV PSNR/SSIM, while M4 had slightly higher LBV PSNR.Source: §3 · p. 6, Table 3 · p. 7
- Both LBV/HBV models preserved expected contrast and prostate morphology, but M2 sometimes showed sharper mask-conforming lesion boundaries, an observation not assessed by experts.Source: §3 · p. 7
4 Conclusion
The framework supports sequential and joint anatomy-conditioned synthesis of prostate bpMRI, with both strategies shown to be feasible. Direct synthesis of ADC and cHBV outperformed deriving them from synthetic LBV and HBV, but diagnostic preservation and clinical value still require validation.
4 Conclusion
- Direct ADC and cHBV synthesis achieved better distribution-based performance than deriving them from synthetic LBV and HBV.The authors suggest that errors introduced during diffusion-image synthesis may propagate to derived representations.Source: §4 · p. 9
- The conditional 3D DDPM framework supports both sequential and joint generation of T2w and diffusion modalities from anatomical masks.Joint synthesis generates complete bpMRI examinations within a single model.Source: §4 · p. 9
- Further validation is needed to determine whether generated images preserve diagnostically relevant characteristics and provide clinical value.Planned evaluation includes expert radiological assessment and downstream classification, detection, and segmentation tasks.Source: §4 · p. 9