Targeted Deep Regression - Initial Applications in Whole-Body Water-Fat MRI in UK Biobank

Filip Marttala; Taro Langner; Johan Öfverstedt; Nouman Ahmad; Håkan Ahlström; Robin Strand; Joel Kullberg

cs.CVOct 26, 2026

Targeted Deep Regression - Initial Applications in Whole-Body Water-Fat MRI in UK Biobank

Authors: Filip Marttala, Taro Langner, Johan Öfverstedt, Nouman Ahmad, Håkan Ahlström, Robin Strand, Joel Kullberg

Organizations: Radiology, Department of Surgical Sciences, Uppsala University, 752 37, Uppsala, Sweden · Department of Information Technology, Uppsala University, 751 85, Uppsala, Sweden · Antaros Medical AB, BioVenture Hub, 431 53, Mölndal, Sweden

Abstract

UK Biobank is a British clinical study which, in addition to large amounts of medical data collected, plans to include detailed Magnetic Resonance Imaging (MRI) examina- tions from 100,000 participants. The image data can be analyzed for example by auto- mated analysis with neural networks for image segmentation or classification to expand the collection of image-derived phenotypes for medical research. In this work, we pro- pose, present and evaluate a methodology that uses image-based regression to estimate numerical properties from the collected neck-to-knee water-fat MRI scans. It applies a ResNet50-style network to 2D multi-channel collages that represent targeted and study- specific anatomical sub-regions from the MRI scan. The targeted sub-region sampling is achieved by propagating manually defined subregions from a reference scan to the other >30,000 analyzed scans using image registration. The resulting collages can be compact while retaining high-resolution, for example by providing multiple slices of kidney tissue for a kidney volume regression network. This approach also enables group-wise saliency analysis that highlight image and anatomical regions used by the network for the pre- dictions. The method was applied for the regression endpoints of kidney volume, heart volume, and age. Predictive performance was found between R2 = 0.731-0.984 and the group-wise saliency maps were found to highlight medically relevant regions. We con- clude that the method can be used to standardize and restrict the information presented to a deep regression analysis pipeline, increasing information density and opening up a wide range of future medical inference studies.

Figures & tables

Figure 1: Illustration of the image analysis pipeline used for the targeted deep regression, with associated cohort saliency analysis that enables interpretability. Template scans (one male and one female) from the 3D water-fat MRI images in UK biobank are first selected. The targeted subsampling of the image data is manually defined using these. Image landmarks, subvolumes and projection are selected and used to target specific anatomical regions of the UK Biobank images and to convert the 3D image data into a 2D representation (e.g. axial mean intensity projection of a subvolume centered on the heart). Deformable image registration is used to transfer the manually selected landmarks and subvolumes into all individual scans. This allows targeted subsampling of the 3D data into a 2D representation from all scans. The extracted multi-channel 2D images are combined into a multi-channel collage for every scan. These collages are used as input to a regression network that is trained to predict a reference value of interest from each individual collage (e.g. the age of a subject). The trained regression networks are furthermore analyzed by performing saliency analysis for each performed inference. This gives an individual saliency result, i.e. a heatmap, for each collage, of what regions in a collage important for the performed task. By use of the image registration results these individual saliency results can be combined and aggregated into a cohort saliency map that shows a cohort-wise result of what regions that are important for the predictions (e.g. what regions are found most important for the age prediction task).

Panel of Figure 2 — Figure 2: An example of a heart collage illustrating the three image channels used (a, b, and c). In this example three projections are used in each channel. Axial (top left), coronal (bottom left), sagittal (bottom right) as well as one sagittal slice (top right) defined using a landmark in the (Aorta).

Figure 3: Examples of water channels of the collages used in the experiments, for kidney volume (far left), left ventricular end diastolic volume (LV-EDV) of the heart (center left), and age from images targeted on the heart using sagittal slices (center right) and axial slices (far right). Note that the packing algorithm used results in a collage with the different image in no particular order. The numbers are not part of the collage image data. They are added to simplify description of the different sub images and referred to as (x) in the text.

Figure 4: STROBE flowchart showing how the participants for experiments were selected

Target	N	Unit	MAE	MAPE	r	$R^{2}$	LoA
Kidney volume	14,182	mL	6.761	0.0277	0.984	0.967	(-18.84 to 19.37)
Heart LV-EDV	14,240	mL	12.324	0.0981	0.856	0.731	(-32.41 to 33.15)
Age sagittal slices	31,164	years	2.832	0.0448	0.880	0.773	(-6.99 to 7.06)
Age axial slices	31,165	years	2.759	0.0437	0.887	0.790	(-6.92 to 6.72)

Table 1: Summary of evaluation metrics from all four experiments, listing number of included subjects (N), mean absolute error (MAE), mean absolute percentage error (MAPE), Pearson’s coefficient of correlation (r), the coefficient of determination (

R^{2}

) and the 95% Limits of Agreement (LoA) from the 10-fold cross-validation.

Panel of Figure 5 — Figure 5: Kidney volume prediction results for males 5(a) and females 5(b) , with plotted numerical results 5(c) .

Panel of Figure 6 — Figure 6: Heart LV-EDV prediction results for males 6(a) and females 6(b) , with plotted numerical results 6(c) .

Panel of Figure 7 — Figure 7: Age (sagittal) prediction results for males 7(a) and females 7(b) , with plotted numerical results 7(c) .

Panel of Figure 8 — Figure 8: Age (axial) prediction results for males 8(a) and females 8(b) , with plotted numerical results 8(c) .

Medical Image Analysis

Medical image analysis uses computational methods to interpret clinical images, measure anatomy and disease, and support diagnosis and treatment.

See all 291 papers

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study
Aug 23, 2026Nabira RashidMedical Image Registration Deformable Image Registration
Carnegie Mellon University · Harvard Medical School
Population-Scale Segmentation of Penile Tissue in DIXON MRI using Deep Learning for Quantitative Phenotyping in Male Reproductive Health
Jul 2, 2026Jan Ernsting, Gunnar Paul Kordes, Nils Johannaber +5Healthcare Magnetic Resonance Imaging
University of Münster Germany
SpikeReg: Energy-Efficient 3D Deformable Medical Image Registration with Spiking Neural Networks
May 24, 2026Ali Mikaeili Barzili, Behzad Moshiri, Hamid Azadegan +1U-Net Medical Image Registration
School of Electrical and Computer Engineering, College of Engineering, University of Tehran, Tehran 1439957131, Iran · Max Planck Institute for Brain Research, 60438 Frankfurt am Main, Germany · School of Computer Engineering, Iran University of Science and Technology (IUST), Tehran 16846-13114, Iran

Aug 23, 2026J/K move · Enter open · S save

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

Nabira Rashid

Carnegie Mellon University · Harvard Medical School

Medical Image Registration Deformable Image Registration

Open paper AskPDF arXiv

Deep learning registration methods routinely stack two kinds of enhancement on a base network: architectural additions such as affine pre-alignment stages, and training-objective additions such as regularization losses. Papers tend to adopt both at once, so it is unclear which is doing the work. I ran a controlled ablation to separate them. Using the OASIS brain MRI dataset (394 training subjects, 20 test subjects), I trained four variants of the same registration pipeline: a baseline 3D U-Net with basic similarity losses, the same U-Net with a full regularization suite, an affine-plus-deformable architecture with basic losses, and the affine architecture with the full suite. I evaluated registration accuracy (MSE, NCC, SSIM), deformation quality (Jacobian determinant preservation, displacement statistics, an anatomical plausibility score), and computational cost. Regularization alone accounted for most of the gain: a 21.3% relative gain on the MSE-improvement metric (1.78% to 2.16%, P<.001) and a 21.8% relative gain in NCC improvement, while cutting maximum deformation from 53.1 to 0.51 units, a 99.0% reduction, at essentially no computational cost (-0.06% inference time). The combined model produced the largest accuracy gain, 25.8% (1.78% to 2.24%), and raised anatomical plausibility from 0.596 to 0.930, at a moderate +9.8% inference-time cost. Gradient correlation rose from 0.742 at baseline to 0.980 for the fully enhanced model. All enhanced variants reached sub-voxel accuracy under plausible deformation constraints. Regularization losses are the primary driver in this setting, delivering the accuracy gains and almost all of the deformation control for free at inference time, while the affine architecture adds a smaller complementary benefit at acceptable cost. The 99% reduction in unrealistic deformations addresses a known barrier to clinical deployment.

Abstract

Medical Image Analysis

Explore similar work

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

Population-Scale Segmentation of Penile Tissue in DIXON MRI using Deep Learning for Quantitative Phenotyping in Male Reproductive Health

SpikeReg: Energy-Efficient 3D Deformable Medical Image Registration with Spiking Neural Networks