MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models

A recent paper on arXiv (2608.10291) explores the potential of lossy compression for training 3D generative models using brain tumor MRI data. Titled 'MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models', the research tackles the high storage and I/O demands of extensive multi-modal MRI datasets, which hinder training on standard infrastructure. Although lossy compression is recognized for maintaining accuracy in discriminative segmentation networks, its impact on generative models, which require learning the complete data distribution, remains unexamined. The authors utilized JPEG2000 or a near-lossless JPEG-LS pipeline to compress each 3D volume and subsequently trained a Wavelet Flow Matching model based on BraTS image sequences (T1n, T1c, T2, T2f). Remarkably, at a 20:1 compression ratio, the synthesis quality matched that of models trained with uncompressed data, indicating that conventional image codecs can efficiently compress complex brain tumor MRI while retaining fidelity for generative tasks. This paper is accessible on arXiv and was noted as a cross-type submission.