The Brain Tumor Sequence Registration Challenge: Establishing Correspondence between Pre-Operative and Follow-up MRI scans of diffuse glioma patients

Bhakti Baheti, Diana Waldmannstetter, Satrajit Chakrabarty, Hamed Akbari, Michel Bilello, Benedikt Wiestler, Julian Schwarting, Evan Calabrese, Jeffrey D. Rudie, Syed Amaan Abidi, Mina Mousa, Javier E. Villanueva-Meyer, Daniel S. Marcus, Christos Davatzikos, Aristeidis Sotiras, Bjoern H. Menze, Spyridon Bakas

Abstract

Registration of longitudinal brain Magnetic Resonance Imaging (MRI) scans containing pathologies is challenging due to tissue appearance changes, and still an unsolved problem. This paper describes the first Brain Tumor Sequence Registration (BraTS-Reg) challenge, focusing on estimating correspondences between pre-operative and follow-up scans of the same patient diagnosed with a brain diffuse glioma. The BraTS-Reg challenge intends to establish a public benchmark environment for deformable registration algorithms. The associated dataset comprises de-identified multi-institutional multi-parametric MRI (mpMRI) data, curated for each scan's size and resolution, according to a common anatomical template. Clinical experts have generated extensive annotations of landmarks points within the scans, descriptive of distinct anatomical locations across the temporal domain. The training data along with these ground truth annotations will be released to participants to design and develop their registration algorithms, whereas the annotations for the validation and the testing data will be withheld by the organizers and used to evaluate the containerized algorithms of the participants. Each submitted algorithm will be quantitatively evaluated using several metrics, such as the Median Absolute Error (MAE), Robustness, and the Jacobian determinant.

Introduction: The Registration Challenge

Medical image registration—aligning two images into a single coordinate system—is a cornerstone of radiology. However, registering a pre-operative scan (containing a large tumor) to a follow-up scan (post-surgery) remains a significant hurdle for standard algorithms. The BraTS-Reg challenge identifies three primary "pain points" in this process:

  1. Mass Effect: Large tumors physically displace healthy brain tissue, creating complex, non-linear deformations.
  2. Missing Correspondences: Surgical resection creates a cavity. The tissue present in the pre-operative scan simply does not exist in the follow-up, violating the fundamental assumption of many algorithms that every point in the source image has a corresponding point in the target.
  3. Intensity Inconsistency: Post-surgical changes, such as scarring or radiation effects, alter tissue appearance across different MRI modalities.
Comparison of pre-operative and post-operative MRI scans showing tumor, edema, and the resulting surgical cavity.

Methods and Dataset

The challenge utilized a multi-institutional dataset of 259 diffuse glioma patients. Each case included four MRI modalities (T1, T1-CE, T2, and FLAIR). To provide a "ground truth," clinical experts manually placed between 6 and 50 landmark points per case on anatomically distinct locations, such as blood vessel bifurcations.

Successful algorithmic strategies among the 15 participating teams typically included:

  • Deep Learning Backbones: Most teams utilized CNN-based frameworks like VoxelMorph or Transformer-based architectures like TransMorph to predict displacement fields.
  • Pre-alignment: Initial affine (linear) registration was critical to bring the scans into a common space before applying deformable models.
  • Test-Time Optimization: The top-performing teams refined their model weights for each specific pair of scans during inference to minimize registration error, a technique that proved to be a significant performance differentiator.

Key Insights and Results

A critical component of the challenge was establishing a human baseline. By measuring inter-rater variability (the disagreement between two experts), the organizers found a median error of 1.41mm. This served as the gold standard for algorithmic performance.

Analysis of inter-rater variability (a) and its correlation with distance to the tumor (b).

Analysis of the results revealed that registration error is significantly higher for landmarks located within 30mm of the tumor core. This confirms that the "mass effect" and the immediate vicinity of the pathology are the primary sources of algorithmic failure.

Despite these challenges, the top-ranked method (Team cwmok) achieved a Median Euclidean Error (MEE) of 1.60mm on the hidden test set. Remarkably, the leading algorithms reached an accuracy equal to or better than human inter-rater variability for approximately 60% of the evaluated landmarks.

Conclusion and Future Work

The BraTS-Reg challenge demonstrates that while deep learning models are approaching human-level precision in longitudinal brain registration, significant work remains. Most algorithms still struggle with the "black hole" of the resection cavity. Future research must focus on "cost-function masking"—explicitly instructing algorithms to ignore areas of missing tissue—and increasing landmark density to better evaluate deformations within the brain parenchyma. For clinical adoption, these tools must also evolve to be more computationally efficient for real-time surgical applications.

Reference: Baheti, B., et al. "The Brain Tumor Sequence Registration (BraTS-Reg) Challenge: Establishing Correspondence Between Pre-Operative and Follow-up MRI Scans of Diffuse Glioma Patients." arXiv:2112.06979v2, 2024.

Summary updated

Abstract

Registration of longitudinal brain Magnetic Resonance Imaging (MRI) scans containing pathologies is challenging due to tissue appearance changes, and still an unsolved problem. This paper describes the first Brain Tumor Sequence Registration (BraTS-Reg) challenge, focusing on estimating correspondences between pre-operative and follow-up scans of the same patient diagnosed with a brain diffuse glioma. The BraTS-Reg challenge intends to establish a public benchmark environment for deformable registration algorithms. The associated dataset comprises de-identified multi-institutional multi-parametric MRI (mpMRI) data, curated for each scan's size and resolution, according to a common anatomical template. Clinical experts have generated extensive annotations of landmarks points within the scans, descriptive of distinct anatomical locations across the temporal domain. The training data along with these ground truth annotations will be released to participants to design and develop their registration algorithms, whereas the annotations for the validation and the testing data will be withheld by the organizers and used to evaluate the containerized algorithms of the participants. Each submitted algorithm will be quantitatively evaluated using several metrics, such as the Median Absolute Error (MAE), Robustness, and the Jacobian determinant.
BibTeX
@misc{baheti2021brain,
  author = {Bhakti Baheti and Diana Waldmannstetter and Satrajit Chakrabarty and Hamed Akbari and Michel Bilello and Benedikt Wiestler and Julian Schwarting and Evan Calabrese and Jeffrey D. Rudie and Syed Amaan Abidi and Mina Mousa and Javier E. Villanueva-Meyer and Daniel S. Marcus and Christos Davatzikos and Aristeidis Sotiras and Bjoern H. Menze and Spyridon Bakas},
  title = {{The Brain Tumor Sequence Registration Challenge: Establishing Correspondence between Pre-Operative and Follow-up MRI scans of diffuse glioma patients}},
  year = {2021},
  eprint = {2112.06979},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.06979},
}
Figure