The research, conducted by Jangho Kwon and Kihwan Choi of the Department of Applied Artificial Intelligence at Seoul National University of Science and Technology, introduces a data-driven, multi-contrast windowing method that learns which contrast settings are most useful for a given image analysis task. Rather than relying on hand-picked window widths and levels, the approach trains a neural network module that suggests multiple windows simultaneously, each tuned to the needs of a downstream segmentation model. The result is a pipeline in which the machine not only detects pathology but also reveals, through automatically generated contrast-enhanced images, which parts of the intensity spectrum it considers important for its predictions.

Windowing works by adjusting two parameters: the window width, which controls the dynamic range of displayed intensities, and the window level, which sets the center of that range and therefore the overall brightness. In a CT image, where Hounsfield units span from dense bone to air, a soft-tissue window might compress the display range to roughly 50 to 350 Hounsfield units, making subtle differences within the liver or brain visible. Radiologists have known for decades that this choice is consequential. A 1999 study in Radiology demonstrated that dedicated liver window settings measurably improved the detection of hepatic lesions, and clinical practice has since accumulated an arsenal of preset windows for different organs and pathologies. In magnetic resonance imaging, the challenge is compounded by the multiplicity of pulse sequences, T1-weighted, T2-weighted, and fluid-attenuated inversion recovery images each carry different intensity distributions, and standardization of these scales remains an active research problem in its own right.

When deep learning entered medical imaging, many researchers simply carried over standard clinical windows, or applied simple normalization schemes such as min-max scaling or z-score standardization, without asking whether those choices were optimal for the model. Kwon and Choi argue this is a blind spot. The input transform applied to a network, including windowing, is itself a hyperparameter of the entire system, and a poor choice can obscure precisely the intensity gradients that a segmentation network needs to delineate a tumor’s boundary. Previous efforts have acknowledged the problem: some studies have trained networks on multiple fixed windows and combined their outputs, while others have proposed trainable windowing for specific tasks such as intracranial hemorrhage detection or liver CT segmentation. The new work extends this line of thinking in two significant directions.

The first is the multi-contrast aspect. Instead of committing to a single learned window, the method generates several windowed versions of each input image, each emphasizing a different band of intensities. These multi-contrast images are then fed to subsequent segmentation models, allowing the network to consult complementary views of the same anatomy. The idea echoes how radiologists themselves work, flipping between lung, bone, and soft-tissue windows to build a complete picture. The second contribution is interpretability. Because the learned windows are task-specific, the method can render a contrast-enhanced image that visualizes which windows the downstream model relies on most heavily for its prediction. In other words, the technique produces a kind of window-level attention map, offering clinicians a window into the machine’s decision-making that goes beyond conventional saliency methods such as Grad-CAM.

To achieve this, the authors construct a windowing module that can be inserted into an end-to-end training pipeline and optimized jointly with the segmentation network. The module learns to remap pixel values so that the regions of interest gain contrast at the expense of irrelevant background intensity ranges. The architecture draws on established building blocks from computer vision, including inverted residual structures familiar from MobileNetV2 and squeeze-and-excitation style channel attention, which allow the module to weigh the relative importance of different learned windows dynamically. During training, the whole system is optimized with gradient-based methods so that the windows adapt to whatever the segmentation task demands, whether that is finding a hypodense liver tumor in CT or distinguishing edema from enhancing tumor core in brain MRI.

The experimental evaluation covered three distinct tasks. The first was liver tumor segmentation in CT images, using data drawn from the well-known Liver Tumor Segmentation Benchmark, or LiTS, a widely used community dataset of contrast-enhanced abdominal CT volumes with expert annotations of liver parenchyma and tumors. The second was abdominal organ segmentation in MRI, assessed in the context of the CHAOS combined CT-MR challenge, which tests models on healthy abdominal structures across different modalities. The third was brain tumor segmentation in MRI, a task made notoriously difficult by the heterogeneous intensity signatures of gliomas and the interplay of multiple MRI sequences. Across these benchmarks, the authors compared their multi-contrast windowing against conventional fixed-window preprocessing and against other segmentation backbones, including attention-based U-Net variants, autoencoder-regularized 3D networks, and transformer-based architectures such as Swin UNETR.

The results, according to the study, show consistent gains. Segmentation models that received multi-contrast windowed inputs achieved higher accuracy than the same models fed conventionally windowed images, indicating that the learned windows were indeed capturing intensity information that fixed windows discarded. Just as importantly, the method produced interpretable visual outputs: contrast-enhanced images in which the band of intensities most critical to the model’s decision was emphasized. For a clinician, this means the AI system does not function as an opaque oracle. It effectively communicates, in the visual language of radiology, what it is looking at, an important step for building the trust needed before such systems enter routine diagnostic workflows.

The implications extend beyond the three tasks studied. Segmentation accuracy is a bottleneck for a wide range of clinical applications, from radiation therapy planning, where tumor boundaries determine treatment volumes, to organ-at-risk delineation, surgical navigation, and quantitative imaging biomarkers. If a simple, learnable preprocessing step can meaningfully improve performance without altering the underlying model architecture or requiring additional hardware, it represents an unusually cost-effective upgrade. The method is also modality-agnostic in principle: any imaging pipeline in which the mapping from raw intensity to display value is somewhat arbitrary, whether cone-beam CT, mammography, or microscopy, could in principle benefit from task-specific learned windowing.

There are also subtle scientific insights embedded in the approach. By examining the learned windows across tasks, one can ask whether a network detecting liver tumors converges on windows resembling the clinical liver window, or whether it discovers entirely different intensity bands. The study’s visualization capability makes such questions tractable, potentially informing radiology practice itself: if a machine consistently prefers a certain window for a certain task, that window might reveal contrast relationships that human observers have overlooked, or confirm decades of accumulated radiological wisdom from a new direction.

The work, published online on 13 May 2026 and funded by a research program of Seoul National University of Science and Technology, builds on the authors’ earlier 2020 conference paper on trainable multi-contrast windowing for liver CT segmentation. In the years since, the field has moved toward ever more powerful segmentation architectures, from nnU-Net, a self-configuring framework that dominates many biomedical segmentation leaderboards, to transformer-based models. Yet the preprocessing layer, the humble transformation that determines what the network actually sees, has received comparatively little attention. This study suggests that revisiting that layer with modern deep learning tools can pay dividends that rival changes in architecture.

For the growing community developing AI-based diagnostic tools, the message is clear: the inputs matter as much as the models. A network is only as good as the representation it is fed, and in medical imaging, that representation is shaped long before the first convolutional filter fires. By making windowing itself a learned, task-adaptive, and interpretable component, Kwon and Choi have turned a routine display setting into a source of both accuracy and insight. As deep learning systems move closer to the clinic, techniques like this one, which improve performance while making the machine’s reasoning visually legible to the radiologists who must ultimately trust it, may prove as important as any architectural breakthrough. The study’s approach of learning multi-contrast windows jointly with segmentation models offers a template that other groups can adopt, extend, and test across new modalities, institutions, and disease targets, bringing the field one step closer to medical AI that both sees better and explains itself.

FAIS is a disorder in which abnormal bony contact between the femoral head-and-neck junction and the acetabulum, the socket of the hip joint, causes pain, cartilage damage, and, in many patients, early-onset osteoarthritis. The most common bony culprit is the so-called cam deformity, a bulge of excess bone at the anterolateral head-neck junction that alters the normally spherical profile of the femoral head. Detecting and quantifying that deformity currently depends on imaging: plain radiographs provide a first look, while CT and magnetic resonance imaging offer the three-dimensional detail needed to characterise the shape and location of the bump and to guide arthroscopic or open surgical correction. But CT exposes young patients, who often make up the FAIS population, to ionising radiation, and MRI is expensive and poorly suited to capturing cortical bone surfaces directly.

The Dalhousie team’s alternative is a freehand 3D ultrasound system, a class of technology in which a conventional two-dimensional ultrasound probe is swept across the skin while its position and orientation in space are tracked continuously. Each B-mode frame, tagged with its pose, becomes a slice in a virtual volume that software can later reconstruct into a three-dimensional representation of the underlying anatomy. What has historically limited this approach for bone imaging is segmentation: bone appears in ultrasound as a bright, often incomplete hyperechoic line with shadowing beneath it, and manually delineating those surfaces across thousands of frames is impractical in a clinical setting. The new system solves this bottleneck with deep learning.

The researchers paired an off-the-shelf optical tracking system with a low-cost point-of-care ultrasound probe. A feature pyramid network, a convolutional neural network architecture originally developed for object detection, was trained to segment bone surfaces from B-mode images. Critically, rather than building a training set from scratch, the team began with a publicly available large-scale dataset of ultrasound bone images, and then refined the model through transfer learning using a smaller set of curated, manually segmented images specific to their application. Transfer learning allows knowledge gained on one large, general dataset to be repurposed for a narrower task with far less labelled data, dramatically reducing the labour of building a clinically deployable segmentation model.

From the segmented bone lines in each tracked frame, deterministic algorithms assembled point clouds of the femoral head and neck surfaces. These were then converted into watertight surface meshes, transforming a collection of partial, overlapping ultrasound sweeps into a coherent three-dimensional model of the bone. The entire pipeline, from probe to mesh, was designed around inexpensive, accessible hardware and open-source components, a deliberate choice that lowers the barrier to adoption compared with specialised, high-cost imaging platforms.

To validate the system, the team compared ultrasound-derived bone surfaces against CT reconstructions in four cadaveric hips and two patient hips. The results were striking: geometric agreement with CT, as well as within-subject repeatability and within-subject reproducibility, all fell in the range of 0.4 to 0.6 millimetres. Those figures matter because the clinically significant bone deformities associated with FAIS measure two millimetres or more. A system that deviates from CT by less than half that threshold can reliably resolve the bumps and asphericities that surgeons need to see. Repeatability and reproducibility matter equally, since a diagnostic tool that returns different answers when the same hip is scanned twice, or scanned by different operators, cannot support confident clinical decisions.

The team also tested the system on the measurement that dominates FAIS assessment: the alpha angle. The alpha angle quantifies the head-neck asphericity by measuring, in a defined plane, the angle between the axis of the femoral neck and the line from the centre of the femoral head to the point where the bony contour departs from a circle. Values above roughly 55 to 60 degrees are conventionally associated with cam impingement, and clinically meaningful differences between measurements exceed five degrees. In the validation experiments, alpha angles derived from the ultrasound reconstructions agreed with those derived from CT to within 0.6 degrees, a margin of error an order of magnitude smaller than the clinically relevant threshold.

The group then moved beyond cadavers and patients already undergoing imaging. Five healthy volunteers underwent scanning to assess feasibility in a live clinical environment, and the system successfully produced three-dimensional bone surface maps of the femoral head in vivo. The demonstration suggests the workflow, sweep the probe, let the network segment the bone, reconstruct the mesh, compute the metrics, is practical for real patients rather than only in controlled laboratory conditions. A supplementary video accompanying the publication illustrates the system in operation.

The implications extend beyond FAIS. Freehand 3D ultrasound has been explored for decades in applications ranging from spinal imaging to foetal biometry and liver interventions, but bone has remained one of the hardest targets because of the physics of ultrasound-tissue interaction. By combining modern semantic segmentation networks, public training datasets, transfer learning, optical tracking, and robust surface reconstruction, the Dalhousie study provides a template that other groups could adapt for osteoarthritis assessment, fracture evaluation, pediatric hip monitoring, and intraoperative guidance, all without radiation. For orthopaedic surgeons, the prospect of quantifying a cam deformity in the clinic, at the point of care, with a probe and a tracker rather than a CT scanner, represents a meaningful shift in diagnostic economics and accessibility.

The authors are careful to frame the work as validation rather than deployment; the study sample was small, comprising four cadaveric hips, two patient hips, and five volunteers, and broader clinical adoption will require larger multicentre studies, demonstration of performance across diverse patient anatomies and body habitus, and integration into existing orthopaedic workflows. The system’s reliance on optical tracking also imposes practical constraints: the probe must remain within the tracker’s field of view during a sweep, which shapes how operators scan the hip. Nevertheless, the combination of 0.4 to 0.6 millimetre accuracy, 0.6 degree agreement on the alpha angle, and successful in vivo imaging meets or exceeds every performance threshold the authors set out as clinically necessary. As the researchers conclude in the paper, the system as demonstrated is suitable for clinical deployment, and the path from laboratory prototype to a radiation-free, affordable tool for hip morphometry has never looked shorter.

A team at Washington University in St. Louis has developed a method intended to solve that problem. The technique, called NGSE-Corr, allows researchers to compare quantitative imaging methods without relying on a gold standard—the definitive, independently verified value against which other measurements are judged. The approach could help scientists evaluate AI-powered imaging systems, assist physicians in selecting more dependable tools and give regulators a way to assess emerging technologies before they enter widespread clinical use. The findings were reported in IEEE Transactions on Medical Imaging in a study led by Abhinav Jha, an associate professor of biomedical engineering at the McKelvey School of Engineering and of radiology at WashU Medicine’s Mallinckrodt Institute of Radiology.

The need for an alternative to the gold standard is widespread in medicine. Consider the challenge of measuring a cancerous tumor inside a living patient. The actual size, biological activity or treatment response of the tumor cannot always be known with complete certainty. A biopsy may sample only a small portion of a complex mass, while a highly accurate reference measurement may require invasive procedures, surgery or extensive follow-up. In other situations, the desired quantity may not be directly observable at all. Researchers may therefore have several imaging tools that measure the same clinical property but no absolute benchmark that reveals which one is closest to reality.

Without such a benchmark, conventional validation can become difficult. A method may agree with another method without either being accurate, or it may appear inconsistent because of variations in image quality, patient anatomy or scanning conditions. Quantitative imaging systems also contain measurement noise: random fluctuations that arise from the scanner, reconstruction algorithms, biological motion and other sources. When several tools are applied to the same patient or tumor, however, their errors are not necessarily independent. Because the instruments observe the same underlying anatomy and are often affected by common conditions, their fluctuations can be correlated.

That insight is central to NGSE-Corr. The method builds on an earlier mathematical formulation for evaluating quantitative imaging, but modifies it to account explicitly for correlated noise among measurements. In simplified terms, the technique examines how different imaging methods vary across repeated or related observations and uses those patterns to estimate their relative precision. Rather than asking whether a measurement matches a known truth, NGSE-Corr asks which method produces the most dependable information under the same clinical circumstances. This distinction is important because precision—the consistency of a measurement—is often assessable even when absolute accuracy cannot be directly established.

Jha and his collaborators, including first author Yan Liu, tested the method through numerical experiments designed to mimic different measurement conditions. Their results indicated that NGSE-Corr could correctly rank imaging methods according to precision, even when the researchers withheld knowledge of the simulated ground truth. The goal was not simply to identify whether an individual measurement was correct, but to determine which of several competing methods was best suited to the task. That ranking capability could be especially useful in fields where new algorithms appear rapidly and where performance may vary depending on the disease, the imaging protocol or the clinical question.

The researchers then moved from mathematical simulations to a virtual imaging trial. They generated computer-based patients with bone-metastatic castration-resistant prostate cancer who were treated with radium-223, a radioactive therapy used in certain cases of advanced disease. The trial compared three quantitative single-photon emission computed tomography, or SPECT, methods for measuring regional activity uptake. In this context, the imaging systems were being evaluated for how precisely they could quantify the distribution of radioactivity in different regions of the body—information that may help researchers understand treatment delivery and response.

The virtual trial produced striking results. When the methods were evaluated in groups of 50 computer-generated patients, NGSE-Corr correctly ranked the imaging approaches in 91% of the trials without being given the underlying ground truth. It identified the most precise method in 95% of the trials. Increasing the number of virtual patients improved performance further, suggesting that larger clinical datasets may allow the method to distinguish between competing tools with greater confidence. Although virtual trials cannot replace carefully designed studies involving real patients, they provide a controlled environment in which researchers can test whether an evaluation strategy behaves as expected.

The implications extend beyond SPECT or prostate cancer. Quantitative imaging is being used to estimate tumor volume, blood flow, tissue composition, metabolic activity and other clinically relevant properties across radiology and nuclear medicine. AI systems are also being developed to transform images into risk scores and treatment predictions. Such systems may be highly sensitive to the data used to train them, the characteristics of the patient population and the technical details of image acquisition. A method that performs well in one hospital may be less reliable elsewhere. By enabling comparisons without requiring a perfect reference measurement, NGSE-Corr could offer a practical way to monitor and rank these tools across diverse settings.

The researchers say the technique could ultimately strengthen confidence in medical imaging technologies while reducing the cost and time associated with traditional validation. For developers, it may provide an objective framework for comparing algorithms during innovation. For physicians, it could clarify which measurements are most dependable when several tools are available. For regulators, it may offer additional evidence when assessing AI-backed products whose outputs cannot easily be checked against an unquestionable biological truth. The method does not eliminate the need for clinical validation or establish accuracy by itself, but it addresses a major gap: evaluating relative performance when the gold standard is unavailable. As medical images increasingly become sources of numerical data rather than pictures alone, tools such as NGSE-Corr could help ensure that those numbers are precise enough to support decisions that affect patient care.

Subject of Research: A method for evaluating the precision and reliability of quantitative medical imaging tools, including AI-based systems, without a gold standard.

Article Title: NGSE-Corr: A Technique for Objective Clinical Evaluation of Quantitative-Imaging Methods Without a Gold Standard

Web References: Washington University in St. Louis; Abhinav Jha profile; IEEE Transactions on Medical Imaging article: https://doi.org/10.1109/TMI.2026.3707743

References: Y. Liu et al., “NGSE-Corr: A technique for objective clinical evaluation of quantitative-imaging methods without a gold standard,” IEEE Transactions on Medical Imaging. DOI: 10.1109/TMI.2026.3707743

Keywords

Medical imaging, artificial intelligence, quantitative imaging, NGSE-Corr, correlated noise, image analysis, SPECT, prostate cancer, virtual clinical trials, machine learning, radiology, medical technology