UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation
Emmanuelle Bourigault, Amir Jamaludin, Abdullah Hamdi · 2025
Read UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation by Emmanuelle Bourigault, Amir Jamaludin, Abdullah Hamdi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-06908.txt.
Introduction
Emmanuelle Bourigault, Amir Jamaludin, and Abdullah Hamdi release a vast labeled MRI dataset for body organ segmentation in this 2025 paper. Medical AI starves for labeled scans, which are costly and private. Their set draws on the UK Biobank's imaging archive. It holds tens of thousands of 3D scans across seventy-two organs. Over a billion labeled slice masks come with them. Automatic labeling did the heavy lifting at that scale.
A cleaning pipeline with organ-specific filters polished the labels. A hand-labeled subset of three hundred scans checks the quality. Trust in automatic labels is the dataset's make-or-break claim. Models trained on it should generalize across hospitals better. The paper is a dataset release with validation. It serves the medical imaging community.
Catalog record
| CREATOR | Emmanuelle Bourigault, Amir Jamaludin, Abdullah Hamdi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-06908
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation Emmanuelle Bourigault, Amir Jamaludin, Abdullah Hamdi Abstract In medical imaging, the primary challenge is collecting large-scale labeled data due to privacy concerns, logistics, and high labeling costs. In this work, we present the UK Biobank Organs and Bones (UKBOB), the largest labeled dataset of body organs of 51,761 MRI 3D samples (17.9 M 2D images) and a total of more than 1.37 billion 2D segmentation masks of 72 organs based on the UK Biobank MRI dataset. We utilize automatic labeling, introduce an automated label cleaning pipeline with organ-specific filters, and manually annotate a subset of 300 MRIs with 11 abdominal classes to validate the quality (UKBOB-manual). This approach allows for scaling up the dataset collection while maintaining confidence in the labels. We further confirm the validity of the labels by the zero-shot generalization of trained models on the filtered UKBOB to other small labeled datasets from a similar domain ( e.g . abdominal MRI). To further elevate the effect of the noisy labels, we propose a novel Entropy Test-time Adaptation (ETTA) to refine the segmentation output. We use UKBOB to train a foundation model ( Swin-BOB ) for 3D medical image segmentation based on Swin-UNetr, achieving state-of-the-art results in several benchmarks in 3D medical imaging, including BRATS brain MRI tumour challenge (+0.4% improvement), and BTCV abdominal CT scan benchmark (+1.3% improvement). The pre-trained models and the code are available at https://emmanuelleb985.github.io/ukbob , while filtered labels will be made available with the UK Biobank. Figure 1: UKBOB Size and Diversity. Our proposed UK Biobank Organs and Bones (UKBOB) is the largest labeled medical imaging dataset for segmentation, comprising body organs of 51,761 MRI 3D samples (17.9 M 2D images) and a total of more than 1.37 billion 2D masks of 72 organs. Left : we show label examples from UKBOB from axial, coronal, and sagittal views. Right : We show a plot of the size (number of 2D images) and diversity (number of classes) of our UKBOB compared to other medical images datasets. The size of the bubbles indicates 2D image resolution. This new scale in dataset size and diversity should unlock a new wave of applications and methods in the computer vision and medical imaging communities. 1 Introduction The advent of large-scale labeled datasets such as ImageNet [ 62 ] and LAION [ 63 ] has been a cornerstone in the remarkable progress of computer vision, enabling the development of powerful foundation models [ 57 , 60 ] that excel across various tasks. These models benefit immensely from the abundance of labeled data, which allows them to learn rich and generalizable representations. In stark contrast, the medical imaging domain grapples with a significant scarcity of large-scale labeled datasets due to stringent privacy regulations, complex logistics, and the high costs associated with expert annotations [ 61 , 26 ] . This limitation hampers the development of robust and generalizable models for critical tasks such as 3D medical image segmentation, which is essential for disease diagnosis, treatment planning, and patient monitoring. Previous medical image datasets either lack diversity or are too small for generalization [ 15 , 2 , 25 ] . Recognizing the pressing need for extensive and diverse medical imaging datasets, we introduce the UK Biobank Organs and Bones (UKBOB) , the largest labeled segmentation medical imaging dataset to date. Based on the UK Biobank MRI dataset [ 66 ] , UKBOB comprises 51,761 3D MRI scans and over one billion 2D segmentation masks covering 72 organs. This dataset not only surpasses existing medical imaging datasets in scale but also in anatomical diversity, providing an unprecedented resource for training robust and generalizable models (see Figure UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation ). Table 1 highlights the differences in scope and other aspects between the different datasets. To collect the labels of UKBOB, we leverage automatic labeling based on the newly released TotalVibe Segmentator [ 16 ] . However, the automatic labeling of such a vast dataset introduces challenges related to label noise and quality assurance. To tackle this, we propose a novel mechanism for filtering organ labels based on a statistical Specialized Organ Labels Filter (SOLF). We also collect manual labels from 300 MRIs for 11 abdominal organs acting as validation (UKBOB-manual). We further account for noisy labels and dynamically refine the segmentation based on the model’s confidence using a novel Entropy Test-Time Adaptation (ETTA). These approaches ensure high-quality labels and enhances the model’s robustness. We validate the validity of the labels by demonstrating zero-shot generalization of the trained models on the filtered UKBOB dataset to other datasets from similar domains, such as the AMOS abdomen MRI dataset [ 30 ] and the BTCV abdomen CT dataset [ 15 ] . Leveraging the extensive UKBOB dataset, we train Swin-BOB , a foundation model for 3D medical image segmentation based on Swin-UNetr [ 21 ] . Our model achieves state-of-the-art performance on several benchmarks in 3D medical imaging, including the BRATS brain tumor MRI challenge [ 2 ] and the BTCV abdominal CT scan benchmark [ 15 ] . Our contributions can be summarized as follows: Contributions: (i) We introduce UK Biobank Organs and Bones (UKBOB), the largest labeled dataset of organs, consisting of 51,761 MRI 3D samples and a total of 1.37 billion 2D segmentation masks of 72 organs based on the UK Biobank MRI dataset. (ii) We leverage automatic mechanisms for cleaning and filtering the labels based on body statistics and specialized organ filter, allowing for high-quality scale-up of the labels. The collected labels are validated by a subset of 300 manually annotated labels of 11 abdominal organs. (iii) To further elevate the effect of the noisy labels, we propose a novel Entropy Test-time Adaptation (ETTA) to refine the segmentation outputs. (iv) We train Swin-BOB , a foundation model for 3D medical image segmentation based on Swin-UNetr network [ 21 ] , achieving state-of-the-art results on standard benchmarks in 3D medical imaging. Medical Imaging Segmentation Datasets Attribute BRATS [ 2 ] BTCV [ 15 ] MIMIC-CXR [ 31 ] Abd.Atlas [ 40 ] Total Segmentator [ 72 ] UKBOB (ours) Number of Classes 3 12 1 25 104 72 Number of 3D Samples 1,470 50 N/A 20,460 1,204 51,761 Total Number of 2D Images 227,850 5,000 377,110 673,000 400,000 17,902,080 Number of 2D Label Masks 581,715 425,000 N/A 16,825,000 5,800,000 1,378,913,040 Number of Patients 1,470 50 227,835 N/A 1,204 50,000 Meta Information N/A N/A Text Reports N/A N/A Bone Density + Fat % Scope Brain Abdomen Chest Abdomen Full-Body Full-Body Modality MRI CT X-rays CT CT MRI Specialty Tumour Organs COVID Organs Organs/Bones Organs/Bones 2D Image Resolution (axial) 240\times 240 314\times 214 2500\times 3056 280\times 280 512\times 512 224\times 174 Table 1: Comparison of Different Medical Imaging Segmentation Datasets . We compare our proposed UKBOB to other well-known medical image segmentation datasets in terms of scope, size, and modality. 2 Related Work 3D Segmentation in Medical Imaging. Advancements in deep learning have significantly influenced 3D data processing, leading to various approaches such as point-based methods [ 54 , 55 ] , voxel-based methods [ 46 , 11 ] , and view-based