Abstract
Developing 3D generative models for medical MRI requires spatially consistent, intensity-normalised, and artefact-free training data.
Multi-scanner cohorts introduce two distinct sources of variability: spectral attenuation from interpolation during resampling, and radiofrequency (RF) inhomogeneity from scanner-specific field non-uniformity. We address both through an explicit, standardised preprocessing protocol, which we term a Data Engineering Contract, and its accompanying pipeline implementation applied to 6,680 multicentre T2-weighted prostate MRI volumes from biopsy-confirmed cases. The pipeline reduces inter-volume variance across spatial, geometric, and depth axes, with each parameter derived from prostate MRI acquisition conventions. We provide a large-scale,
training-free characterisation via two analyses: an interpolation kernel audit using 3D radial power spectral density, and a paired within-cohort N4 comparison stratified by vendor and field strength. These results demonstrate that preprocessing decisions are not implementation details: they introduce measurable, systematic variation into the data presented to the generative model. Without an explicit, validated preprocessing protocol, no claim about generative model performance can be unambiguously attributed to architectural choices rather than data engineering decisions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_017.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=cgJWkN3bkY
BibTex
@InProceedings{GioKon_The_MICCAISAT2026,
author = { Giouroukou, Konstantina AND Kalliatakis, Grigorios AND Dimitriadis, Avtantil AND Tsiknakis, Manolis AND Klontzas, Michail E. AND Marias, Kostas},
title = { { The Data Engineering Contract: A Formally Specified Pipeline for Generative Prostate MRI } },
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}
}