3D Transport-based Morphometry (3D-TBM) for medical image analysis

AI summary

Overview

Research area: Medical image analysis and computational anatomy, combining optimal mass transport theory with 3D brain imaging. The paper is a software/tool contribution published under arXiv:2602.07260v1 [cs.CV] (06 Feb 2026), with keywords listed as Medical Image Analysis, Transport-Based Morphometry (TBM), Optimal Transport, and 3D Brain Image.

Technical level: Intermediate. The underlying mathematics (optimal transport, Wasserstein barycenters, penalized discriminant analysis) is advanced, but the paper is written as a tool-and-tutorial paper that hides that math behind simple Python interfaces, so a reader with basic medical imaging or machine learning background can follow it.

Scope: The paper introduces 3D-TBM, a Python tool distributed through PyTransKit for running transport-based morphometry on 3D medical images, and demonstrates it on age classification and age regression using white matter from the IXI brain MRI dataset.

What This Paper Is About

Transport-Based Morphometry (TBM) embeds medical images into a mathematical "transport domain" where morphological differences become easier to analyze linearly, and because the transform is invertible, analysis results can be mapped back onto the original image so clinicians can see which anatomy drove a model's output. Until now, no user-friendly software existed to run this for 3D images — widely used tools such as ANTs, MONAI, FreeSurfer, SPM, and AFNI do not support TBM, and general optimal transport libraries like POT have no TBM-specific functions. The goal of this paper is to provide 3D-TBM, a complete, accessible pipeline covering preprocessing support, Linear Optimal Transport (LOT) embedding, statistical analysis, and visualization.

Key Contributions

1. A complete Python pipeline for 3D transport-based morphometry. 3D-TBM covers data loading, computation of the Linear Optimal Transport (LOT) embedding, statistical analysis, and inverse visualization, with the source code publicly available through PyTransKit.

2. Encapsulated modules with adjustable parameters. Each stage is exposed through a dedicated interface — including load_data for preprocessing/loading, TBM3D_forward and process_features for LOT computation, SimplePCA for dimensionality reduction, PLDA for discriminant analysis, CCA for correlation analysis, and visualization functions (visualize_result, visualize_pca, inverse_visualiztion, visual_roc, show_3d_arrays). A parallel computing option (parallel=True, num_workers) is provided because LOT computation on large 3D images can be time-consuming.

3. Worked examples and tutorials on the IXI dataset. The paper walks through an age-analysis experiment on IXI white matter — classifying young adults (under 35 years of age) versus older adults (over 60 years of age), and predicting age by regression — with code listings for every step.

4. A classification benchmark. A comparison of PLDA, Nearest Subspace, and Local Nearest Subspace classifiers on 330 IXI samples across 20 random train/test splits.

Main Findings

• Interpretability is the central payoff. Because the optimal transport map is invertible, principal components from PCA and discriminant directions from PLDA can be mapped back to image space using the inverse transform, letting users see the anatomical changes associated with a model's output rather than only a numeric score.

• Age-related ventricle enlargement was recovered from white matter. In the PLDA visualization on IXI white matter, samples further to the left tended to correspond to younger patients and those on the right to older patients, and the authors observed that ventricle size increases with age — a result they state is consistent with previous findings.

• CCA gave a consistent result. The age analysis via CCA and linear regression produced conclusions similar to PLDA, namely that ventricular size increases with age.

• Nearest Subspace outperformed PLDA on the classification task. Results on 330 IXI samples with a 200/130 train/test split repeated 20 times: PLDA achieved accuracy 0.78±0.026 and balanced accuracy 0.777±0.028; Nearest Subspace achieved accuracy 0.846±0.033 and balanced accuracy 0.84±0.034; Local Nearest Subspace achieved accuracy 0.82±0.042 and balanced accuracy 0.822±0.041. The paper describes the improvement over the linear classifier as an increase in accuracy from 0.78 to 0.846.

• Local Nearest Subspace performed less favorably, which the authors suggest may indicate the task is relatively simple and that the data distribution in the transport domain does not exhibit significant complexity.

• Reference choice may affect accuracy. The summary states that the experiment compared classification performance under different choices of references and classifiers, and that using the intrinsic mean as the reference may improve predictive accuracy. The numeric table reported in the paper lists only the three classifiers, not a per-reference breakdown.

• The intrinsic mean sharpens the template. Computing the intrinsic mean (Wasserstein barycenter) requires the LOT embedding and inverse visualization to be completed first, and iteratively updating it via the mean optimal transport map progressively sharpens the template.

Methodology in Plain English

The workflow has four stages, shown in the paper's Figure 1.

Stage 1 — Prepare the data. The user registers, segments, crops, and centers their 3D images; 3D-TBM deliberately does not implement these because they depend on the anatomy and task. Centering is described as mandatory so all images align around a consistent center. Preprocessed images are stored as NumPy arrays and listed with their labels in a CSV file containing columns such as image_path and label. For the IXI demonstration, all preprocessing (registration, white matter segmentation, centering) was done in SPM and the results converted to NumPy.

Stage 2 — Embed images into the transport domain. Images are first normalized so that they sum to 1. A reference image is then required: either the user supplies one, or the mean of the training images is used. An optimal transport map is computed that morphs the reference into each target image under minimal transport cost, subject to a mass-preservation constraint. This is solved with multiscale accelerated gradient descent. The resulting maps are converted into an N×D array via (f*−I_d)·√I_0 by process_features(). Results can be saved to disk to avoid recomputation.

Stage 3 — Analyze. Because each transport map for an h×w×d image has dimensionality 3·h·w·d, dimensionality reduction comes first. 3D-TBM uses PCA (SimplePCA, with an example retaining components explaining 96% of variance, a threshold that is adjustable). On top of that, users can run PLDA for classification — where a regularization parameter α trades off between LDA (lower values) and PCA (higher values), and calculate_alpha() sets α automatically so small changes in α do not substantially alter the mapping — or CCA plus linear regression for correlation and regression tasks. Evaluation uses a confusion matrix, accuracy, balanced accuracy, and AUROC.

Stage 4 — Map results back to images. Sampled points along a principal component or discriminant direction are turned back into deformation fields and inverted into 3D images. For PCA, f_α = f̄ + αϕ_i with α set to −2σ, −1σ, 0, σ, 2σ. For PLDA, f_α = f̄ + αwᵀB with the same α values. Geodesics between an input image and the reference are visualized with f_α = (1−α)·I_d + α·f*_i, where α=0 gives the input and α=1 gives the reference.

Why This Matters

This work lowers the barrier to entry for a technique that has already shown clinical promise but lacked accessible software. Because TBM keeps a link between statistical results and physical anatomy, it produces findings that a clinician can inspect spatially rather than treat as a black-box number — an important property for clinical adoption and for medical AI more broadly.

Real-world applications implied by the paper and its cited prior work:

• Stroke care: prediction of hematoma expansion, where hematoma regions are segmented and cropped before analysis.
• Healthy aging and fitness: studying brain tissue distribution changes associated with cardiorespiratory fitness, and classifying or predicting age from white matter.
• Neurodevelopmental conditions: identification of autism-related endophenotypes via white matter segmentation-based classification.
• Methodological reuse in clinical imaging studies generally: any study where a user wants to classify or regress on 3D morphology while retaining spatial interpretability.

Industry relevance: the tool is distributed through PyTransKit under a CC BY-NC-SA 4.0 license, integrates naturally with the Python scientific stack (NumPy, scikit-learn), and slots into existing research workflows rather than replacing registration, segmentation, or deep learning toolchains. The paper notes that users can additionally apply scikit-learn tools such as support vector machines, k-nearest neighbors, random forests, or deep learning to the transport-domain features. The funding acknowledgments — ONR (N000142212505) and NIH (GM130825, U54-CA274499) — indicate defense- and health-agency interest in the capability.

Future Directions

• Closing the reference-selection gap. The paper claims that using the intrinsic mean as the reference may improve predictive accuracy, but the reported numeric table covers only classifiers. A systematic, quantified comparison of reference choices (mean image versus iteratively refined intrinsic mean/Wasserstein barycenter) is the natural next step.

• Extending the benchmark. Only three classifiers were compared on a single binary age task from IXI with 330 samples. Whether the Nearest Subspace advantage over PLDA holds on other anatomies, other datasets, and multi-class or regression settings is untested here.

• Scaling and runtime. The paper states that LOT computation on relatively large 3D medical images can be time-consuming and offers parallel execution as mitigation, but no runtime or memory measurements are reported, leaving scaling behavior on large cohorts an open question.

• Broader modality and task coverage. All demonstrations are on structural brain MRI from IXI (which includes T1, T2, and PD modalities), and the examples focus on white matter. Whether the pipeline transfers to other organs, other imaging modalities, and other clinical endpoints is not addressed.

Target Audience

Clinical imaging researchers and neuroscientists who want to apply morphological analysis to 3D medical images without reimplementing optimal transport mathematics; biomedical engineering and computer vision researchers studying transport-based representations; and software-minded practitioners who need an interpretable alternative or complement to black-box deep learning models. The paper is written to onboard newcomers — it includes a small-dataset tutorial and step-by-step code — so readers with basic Python and medical imaging familiarity will get the most from it, while readers seeking theoretical proofs of the underlying transport mathematics should look to the referenced prior work instead.

Authors’ abstract

Transport-Based Morphometry (TBM) has emerged as a new framework for 3D medical image analysis. By embedding images into a transport domain via invertible transformations, TBM facilitates effective classification, regression, and other tasks using transport-domain features. Crucially, the inverse mapping enables the projection of analytic results back into the original image space, allowing researchers to directly interpret clinical features associated with model outputs in a spatially meaningful way. To facilitate broader adoption of TBM in clinical imaging research, we present 3D-TBM, a tool designed for morphological analysis of 3D medical images. The framework includes data preprocessing, computation of optimal transport embeddings, and analytical methods such as visualization of main transport directions, together with techniques for discerning discriminating directions and related analysis methods. We also provide comprehensive documentation and practical tutorials to support researchers interested in applying 3D-TBM in their own medical imaging studies. The source code is publicly available through PyTransKit.