#12 · ML / AI methods

Learned models separate MRI artifacts from anatomy and fuse genomic and imaging features

The common thread across these preprints is explicit modeling of structured variation in medical imaging, whether technical or biological. DisMorph, trained on synthetic data, separates MRI gradient distortion from true anatomical change in longitudinal deformation fields. GeneFuse combines pre-trained genomic language model embeddings with neuroimaging features via FiLM-style modulation and uncertainty-gated fusion. CAN-FLOW applies two-step conditional normalizing flows over diffeomorphic cardiac shape momenta to generate metadata-conditioned biventricular anatomies for virtual cohorts. The dominant theme is conditioning and disentanglement rather than single-cohort prediction, with each method making a specific source of variation explicit in the model. What is new here is the extension beyond brain imaging to cardiac anatomy and to genomic priors, with synthetic training used in at least one case.