Abstract
Fetal brain pathologies, such as hydrocephaly and severe ventriculomegaly, are rare but clinically significant conditions that require accurate prenatal diagnosis. The limited availability of pathological ultrasound images, combined with high variability in imaging due to gestational age, probe angle, and operator technique, poses a major challenge for training robust AI models for diagnostic support tasks.
To address this, we propose a controllable synthesis pipeline for transventricular (TV) fetal brain ultrasound that couples registration‑guided mask with radiomics‑conditioned latent diffusion. First, posterior ventricle and choroid plexus masks from real pathological cases are aligned to normal TV scans using a skull‑anchored affine‑then‑diffeomorphic registration strategy, ensuring plane‑specific anatomical localization. Next, we apply radiomics-conditioned inpainting in a VQ-VAE latent space using a Latent Brownian Bridge Diffusion Model (LBBDM). The model is conditioned on (i) the masked normal image, (ii) fixed intensity prior for key structures, and (iii) radiomics texture features sampled from the training distribution, yielding realistic and controllable pathology appearance. In a downstream evaluation for hydrocephaly image similarity-based retrieval, augmenting real-world training data with our synthetic images increases the expected number of diagnosis-matching references in a top 10 query from 0.35 to 0.7 with a self-supervised ResNet-18 encoder. This work demonstrates that incorporating synthetic data boosts retrieval performance for rare fetal pathologies.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=NkngS1EWNF
BibTex
@InProceedings{ZsaRic_Controllable_MICCAISAT2026,
author = { Zsamboki, Richard AND Czipczer, Vanda AND Olajos-Horvath, Alinka},
title = { { Controllable Synthesis of Pathological Fetal Brain Ultrasound via Registration-Guided, Radiomics-Driven Latent Diffusion } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}