UPCommons. Portal del coneixement obert de la UPC
Publicación
Contribución de congreso (ConferenceOutput)
Anatomically guided 3D diffusion models for synthetic prostate MRI generation
Automated analysis of prostate MRI has the potential to improve cancer diagnosis, but its development is limited by the scarcity of large, annotated datasets. This limitation stems not only from the laborintensive nature of expert annotation but also from strict patient privacy constraints that restrict data sharing. Diffusion-based generative models show promise for data augmentation, but often lack anatomical awareness, limiting their ability to produce realistic lesions and fully coherent anatomy. To address this, we propose an anatomically guided 3D diffusion model that synthesizes T2-weighted prostate MRI conditioned on region masks (prostate zones and lesions). In addition, our method leverages FiLM-based label conditioning to enable class-specific generation for clinically significant prostate cancer. We evaluate our approach using both standard image synthesis metrics and performance of downstream region segmentation tasks. In particular, we show that nnU-Net models trained on synthetic and hybrid datasets match the performance of models trained on the full set of real annotations, despite using only 17% of the original manual annotations. Furthermore, radiomic feature analysis demonstrates that descriptors from synthetic images closely align with those from real MRI, confirming structural plausibility. These results highlight the potential of anatomically conditioned diffusion models to generate high-quality synthetic data, reduce the annotation burden, and provide a scalable solution to develop deep learning models in prostate MRI., This work was supported by the European project Federated Learning and mUlti-party computation Techniques for prostatE cancer (HORIZON101095382-FLUTE), the Spanish Research Agency (AEI) under project PID2023- 148614OB-I00 research project funded by MICIU/AEI /10.13039/501100011033 and by FEDER, EU, the FPI-Ministerio PRE-2021-098481 grant and the predoctoral program AGAUR-FI ajuts (2025 FI-1 00076) Joan Oró backed by the Secretariat of Universities and Research of the Department of Research and Universities of the Generalitat of Catalonia, as well as the European Social Plus Fund., Peer Reviewed