Learning General-purpose Biomedical Volume Representations using Randomized Synthesis

Authors: Neel Dey, Benjamin Billot, Hallee Wong, Clinton Wang, Mengwei Ren, Ellen Grant, Adrian Dalca, Polina Golland

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments demonstrate that the resulting features and weights enable broad generalization on the key biomedical tasks of 3D registration and segmentation across several diverse datasets. We achieve state-of-the-art unsupervised multimodality image registration by simply using the network s approximately appearance invariant and pose equivariant representations (Fig. 1) to drive existing registration solvers. The proposed network can also be used as an off-the-shelf dataset-agnostic initialization for finetuning on any voxel-level task. Specifically, we demonstrate strong few-shot segmentation performance across several highly variable downstream datasets, thereby removing the need for the cumbersome dataset-specific self-supervised pretraining commonly used today.
Researcher Affiliation Academia 1 MIT CSAIL 2 New York University 3 Harvard Medical School EMAIL
Pseudocode Yes Algorithm 1 3D synthetic label map L generation
Open Source Code Yes Code, tutorials, and model weights are available at https://www.neeldey.com/anatomix/.
Open Datasets Yes As templates, we use the freely available 45,000 binary volumes from the Total Segmentator dataset of 104 annotated organs in 1,204 CT volumes (Wasserthal et al., 2023).
Dataset Splits Yes we split L2RAb and MM-WHS into 1/7 and 5/15 validation/testing pairs, respectively, where the validation pair(s) are only used for tuning registration hyperparameters. [...] The dataset splits are in App. B.5.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. It only mentions a "3T Skyra Siemens scanner" in the context of data acquisition for the WUFetal dataset, not for model training or inference.
Software Dependencies Yes Our data engine in Fig. 2 A & B has several components. Here, we describe low-level implementation details. We note that we make extensive use of the MONAI (Cardoso et al., 2022), Torch IO (P erez Garc ıa et al., 2021), and scikit-image (van der Walt et al., 2014) libraries for both label and volume synthesis. [...] Convex Adam is a high-performance multimodality registration solver and we use the b2671f8 commit of the repository4.
Experiment Setup Yes F and Z are pretrained jointly for 600,000 iterations with a batch size of one 1283 label map, each generating two 1283 volumes. We compute the contrastive loss (with temperature τ = 0.33) on 512 randomly sampled indices at each iteration for each decoder layer due to memory limitations. [...] we finetune all layers of each network for a high number of iterations (37,500) with cosine learning rate decay.