Modular Preprocessing Pipeline Standardizes Breast MRI Data for Artificial Intelligence

Researchers have introduced an open-source, modular preprocessing pipeline named bmri-ai-prep, alongside an adapted data formatting standard, designed to prepare complex breast magnetic resonance imaging datasets directly for artificial intelligence and deep-learning workflows. Magnetic resonance imaging is an exceptionally versatile diagnostic modality, capturing three-dimensional volumetric stacks across multiple contrast sequences, dynamic time courses, and diffusion weightings within a single session. However, prevailing deep-learning vision models were originally engineered for two-dimensional natural images. When applied to breast imaging, these networks encounter substantial obstacles because clinical acquisitions require an expansive field of view to suppress imaging artifacts, leaving the target breast tissue occupying only a minor fraction of the total volume. In conventional practice, researchers have frequently fragmented volumetric data into independent two-dimensional slices, downsampled images, or calculated maximum intensity projections to conform to restricted network dimensions. These ad hoc reduction methods inevitably compromise inter-slice geometric coherence and spatial correspondence across sequences, stripping away vital multidimensional information.

The absence of tailored preprocessing tools stems largely from the historical dominance of brain neuroimaging within medical imaging research. Over decades, the academic community developed sophisticated open-source software packages, including FSL, FreeSurfer, SPM, AFNI, ANTs, and DiPy, specifically optimized for cranial anatomy. Compounding this tool discrepancy is a lack of organizational standardization across public breast imaging repositories. While neuroimaging widely adopted the Brain Imaging Data Structure to standardize directory hierarchies and metadata, public breast MRI datasets remain predominantly in raw DICOM format. This format imposes heavy header overhead and protocol variability, forcing artificial intelligence researchers to construct labor-intensive custom data loaders or execute extensive manual restructuring prior to model development.

To establish an organized bridge between raw archives and deep-learning models, the researchers introduced the Breast Imaging Data Structure, designated BRIDS. Adapted from the neuroimaging BIDS standard, BRIDS replaces complex multi-frame DICOM files with streamlined NIfTI image storage, enforcing consistent directory structures and naming conventions optimized for breast imaging protocols. The accompanying bmri-ai-prep pipeline features a fully modular design where discrete preprocessing steps execute independently without rigid inter-stage dependencies. Progression across data types is tracked through simple text files with the extension *.baip.txt, which record the completion of each processing operation. This file-based tracking allows investigators to audit execution, fine-tune parameters for specific institutional protocols, and seamlessly resume processing after unexpected interruptions, effectively eliminating the friction and overhead of adapting brain-oriented tools or drafting bespoke data parsers.

The bmri-ai-prep pipeline incorporates three core computational modules: intra-sequence motion correction, inter-sequence co-registration, and anatomical segmentation. Intra-sequence motion correction operates primarily on dynamic contrast-enhanced sequences, where patient movement across dynamic frames can corrupt kinetic enhancement profiles. In comparative evaluations of deformation algorithms across ANTs, SimpleElastix, and NiftyReg, deformable registration proved crucial for capturing non-rigid tissue movement. NiftyReg’s deformation implementation offered the most favorable balance of performance, flexibility, and speed, requiring a mean runtime of 312 seconds per series compared to 463 seconds for SimpleElastix and 573 seconds for ANTs. Crucially, the researchers observed that relying solely on objective similarity metrics, such as the Fractional Improvement of Structural Similarity, proved misleading. While ANTs yielded the highest FISS score of 0.0061 against 0.0053 for default NiftyReg, visual inspection revealed anatomically implausible warping. Increasing default deformation regularization penalty terms fivefold to tenfold successfully resolved this distortion.

The quantitative performance of the motion correction module was audited across 40 evaluation subjects from the Duke-Breast-Cancer-MRI and Advanced-MRI-Breast-Lesions datasets. Under the adopted NiftyReg settings, the median landmark displacement residual fell from 1.582 millimeters to 1.155 millimeters, achieving a median per-subject reduction of 33.3 percent. Conversely, SimpleElastix degraded alignment, increasing the displacement residual by 34.3 percent. Crucially, the adopted setting preserved the diagnostic time-to-peak enhancement in 95.0 percent of subjects while shifting peak enhancement by a median of negative 4.65 percentage points; SimpleElastix preserved time-to-peak in only 82.5 percent of subjects. Acquisition protocol comparisons confirmed that displacement residual reductions did not differ significantly between field strengths, averaging 30.7 plus or minus 14.5 percent at 1.5 Tesla versus 28.7 plus or minus 20.7 percent at 3.0 Tesla. Time-to-peak was preserved across all 27 subjects at 1.5 Tesla and in 11 of 13 subjects at 3.0 Tesla, demonstrating consistent stabilization.

To enable multi-parametric artificial intelligence analysis combining distinct functional contrasts, the co-registration module aligns all secondary sequences to a pre-contrast dynamic contrast-enhanced reference volume. Clinical protocols routinely incorporate axial T2-weighted and diffusion-weighted imaging alongside dynamic series. Default registration parameters produced severe tissue smearing and structural degradation. Systematic binary search optimization revealed that increasing deformation regularization penalties three- to fivefold established robust cross-modal alignment while preserving anatomical morphology across T1-weighted, T2-weighted, and diffusion sequences. For multi-session longitudinal imaging, the authors implemented a two-stage procedure pairing initial affine registration with subsequent deformation under regularization penalties increased sevenfold to tenfold. In testing quantitative co-registration metrics, including structural similarity indices, ORB feature similarities, and MedicalNet ResNet-50 embedding cosine similarities, none reliably tracked visual quality or regularization changes, confirming that expert visual inspection remains the primary dependable evaluation standard for breast MRI registration.

The segmentation module isolates breast parenchyma, addressing the extreme data sparsity typical of wide-field axial acquisitions that encompass the thoracic cage, heart, lungs, spine, and background air. Quantitative evaluations across 25 manually segmented sessions revealed that breast tissue occupies merely 5.2 plus or minus 2.4 percent of the total acquired volume. The tightest rectangular bounding boxes enclosing both breasts occupy 15.7 plus or minus 5.0 percent, representing a theoretical sixfold reduction in input array size. The bmri-ai-prep pipeline emits conservative bounding boxes with safety margins retaining 26.1 plus or minus 7.4 percent of the volume, providing an approximate fourfold reduction. Concurrently, applying segmentation masks reduces stored file sizes by an order of magnitude, from 100–200 megabytes down to 10–20 megabytes per sequence while improving overall signal-to-noise ratios. This reduction allows volumetric networks to ingest multi-sequence arrays without aggressive downsampling.

Two deep-learning segmentation models were evaluated alongside classical thresholding benchmarks. The investigators developed SegVN, a three-dimensional convolutional neural network based on VNet with leaky ReLU activations, which required downsampling input volumes to 256 by 256 in-plane voxels across 64 slices due to GPU memory constraints. In parallel, they developed SegSF, adapting the SegFormer visual transformer pre-trained on natural images, which processed reformatted slices at 512 by 512 resolution across 112 slices. Across five-fold cross-validation, SegSF converged rapidly, dropping training loss by an order of magnitude within ten epochs, while SegVN required over 500 epochs. A fixed probability threshold of 0.75 for SegVN provided more consistent performance than parametric cluster optimization. SegSF demonstrated superior performance over SegVN, achieving higher Dice Coefficients and lower 95th percentile Hausdorff Distances.

Generalization was rigorously assessed using subject-level five-fold cross-validation across 25 sessions from 23 subjects. Across all 25 held-out sessions, SegSF attained a mean Dice score of 0.897 plus or minus 0.084, compared to 0.957 on training data. Performance varied substantially by source: held-out sessions from the dominant AMBL dataset achieved a 0.931 Dice score, but accuracy dropped to 0.794 on a Duke dataset subject and averaged 0.756 across longitudinal sessions. When compared with nnU-Net, a self-configuring three-dimensional reference framework trained under identical folds, nnU-Net achieved higher accuracy, exceeding SegSF in 23 of 25 sessions (paired t equals 2.90, p equals 0.008) and maintaining Dice scores of 0.941 and 0.954 on non-dominant protocols.

Runtime benchmarks showed that processing a complete multi-sequence case on an NVIDIA RTX 4090 GPU required approximately 10 to 20 minutes, an operational speed comparable to neuroimaging pipelines like fMRIprep. However, key clinical and technical limitations remain. A significant limitation is the exclusion of axillary regions from segmentation masks; because axillary lymph nodes are critical for breast cancer staging and prognosis, omitting them restricts the pipeline’s immediate utility in comprehensive oncologic assessment. Additionally, the evaluation cohort lacked broad multi-vendor representation, with 39 of 40 registration subjects imaged on hardware from a single manufacturer. Furthermore, domain-specific quantitative registration metrics remain unavailable, leaving qualitative visual inspection as the primary evaluation method. By providing bmri-ai-prep and its fine-tuned models under an open-source license alongside BRIDS, the authors deliver a transparent foundation that preserves volumetric coherence, minimizes manual preparation, and supports collaborative development of robust clinical artificial intelligence models.

Subject of Research: Development and evaluation of an open-source preprocessing pipeline and standardized data structure for breast magnetic resonance imaging in artificial intelligence workflows

Article Title: bmri-ai-prep: a preprocessing pipeline for breast MRI data for AI applications

Article References: Xiao, R., & Wang, Y. (2026). bmri-ai-prep: a preprocessing pipeline for breast MRI data for AI applications. Discover Artificial Intelligence, 6(1), Article 1764. https://doi.org/10.1007/s44163-026-02708-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02708-9

Keywords: breast MRI, bmri-ai-prep, deep learning, BRIDS, medical image preprocessing, motion correction, inter-sequence co-registration, anatomical segmentation, NiftyReg, SegFormer, VNet, nnU-Net, dynamic contrast-enhanced MRI, NIfTI, diffusion-weighted imaging, medical image analysis, computer vision in medicine