Skip to content

Medical Imaging Learning — Polina Golland (MIT CSAIL)

ISBI 2026 plenary keynote on statistical methods, MRI reconstruction, and image registration for medical imaging — the methods-side foundation for downstream cancer-MRI workflows even though the keynote itself is method-axis, not oncology-direct.

  • Speaker(s) / Authors: Polina Golland
  • Affiliation(s): MIT CSAIL (Computer Science and Artificial Intelligence Laboratory)
  • Anchor type: keynote
  • Track / day: ISBI 2026 main plenary (day TBD — Thu Apr 9 or Fri Apr 10)
  • IEEE Xplore: TBD (keynote — no Xplore paper)
  • Status: post-meeting

What it does

Golland's keynote covers three of MIT CSAIL's long-running medical imaging axes: (1) accelerated MRI reconstruction via learned priors, (2) deformable image registration with learning-based approaches (the VoxelMorph line), (3) statistical methods for population-level imaging analyses. These methods are the substrate that downstream cancer-imaging workflows depend on — cancer-MRI rapid-reconstruction pipelines, tumor tracking via registration, population-level radiomics studies.

How it works / methodology

The Golland-lab pattern is rigorous statistical-method development plus open-source reference implementations: VoxelMorph (learning-based registration) became the de-facto baseline for deformable registration; learned priors for MRI reconstruction (compressed-sensing-with-learning hybrids) bridge classical inverse-problem theory and modern deep learning. Methods emphasize sample efficiency, interpretability, and demonstrable behavior under distribution shift — qualities that matter when downstream use is clinical.

Headline benchmarks

TBD — keynote-level synthesis; the published Golland-lab work (VoxelMorph, learned MRI reconstruction priors) remains the citation anchor. ISBI 2026 keynote likely surfaced an update or new direction; specific deltas not in program-notes.

How it works

Core idea. Medical imaging benefits from learned methods that are statistically rigorous, sample-efficient, and well-behaved under distribution shift — three properties Golland-lab work optimizes for explicitly. The keynote spans three method axes that share this design philosophy: learned MRI reconstruction priors, deformable image registration (VoxelMorph line), and statistical population-imaging methods.

Inputs / outputs. MRI reconstruction: undersampled k-space → reconstructed image (with learned prior regularizing the inverse problem). Registration: fixed image + moving image → dense deformation field aligning moving to fixed. Population analysis: image cohort + metadata → group-level statistical maps / latent factors.

Key innovation. Hybridizing classical inverse-problem theory (compressed sensing, variational regularization, diffeomorphic registration formulations) with learned components rather than replacing the theoretical scaffold with end-to-end deep learning. VoxelMorph (the most-cited Golland-lab artifact) replaced expensive iterative diffeomorphic registration with a single forward CNN pass while keeping the smoothness / invertibility regularization that classical methods enforced — preserving theoretical guarantees while delivering ~10000x speedup.

Parameters. VoxelMorph: U-Net backbone, ~300K–1M parameters depending on volume size; learned MRI reconstruction priors typically use unrolled-network architectures with shared weights across iterations (~10–20 unrolled steps, parameter counts in the low millions).

Canonical example. Brain MRI cross-subject registration: VoxelMorph maps a moving T1 volume to a fixed T1 template in ~1 second on a GPU vs ~minutes for classical diffeomorphic registration, with Dice on subcortical-structure overlap matching or exceeding ANTs SyN. For cancer-imaging downstream: longitudinal tumor tracking via the same registration backbone, and accelerated cancer-MRI reconstruction for higher-throughput scanning at maintained image quality.

Where it fits in the corpus

  • AACR 2026: indirect — methods feed cancer-MRI workflows but no direct pathology-FM crosswalk
  • MICCAI 2026: strong methods-axis sibling — Golland-lab work consistently appears at MICCAI main proceedings; cousin papers likely Sep 27–Oct 1
  • RSNA 2026: indirect — VoxelMorph and learned MR reconstruction underlie many clinical-translation systems shown at RSNA
  • CVPR 2026: Golland-lab work bridges to CVPR's generative-modeling and inverse-problems tracks

Notes

Keynote-axis page only — no tool artifact directly attached. Watch for any ISBI 2026 main-proceedings paper from Golland-lab students that releases a named model, which would warrant a separate tool page. The keynote serves as a methodological anchor in the vault for "where the ISBI methods axis sits relative to cancer-imaging downstream."