Weight-space programming for auditing clinical decision networks
Humza Nusrat, Omar Nusrat
AI 速览：作者报告前沿语言模型仅凭临床协议和网络权重即可盲审并修复小型放疗决策网络中的部分植入缺陷，但尚不能证明网络正确。
Machine learning models in radiotherapy are commissioned by testing them on cases with known answers, which may not reveal a shortcut that the test cases share with the training data. Frontier language models have recently been reported to read and write the raw weights of small neural networks, an ability called weight-space programming. The purpose of this study was to determine whether such a model, given only a clinical protocol and a network's weights, can audit and repair the network. Networks with 433 to 4,801 parameters were trained to screen head-and-neck planning requests against four rules. Flaws were planted through the training data: confounds that hid each rule, a 3 Gy shift in the cord threshold, and a hidden trigger. Claude Opus 5.5 audited each network blind, without tools, data or code execution, and every claim and repair was tested on about 420,000 requests. The language model named the broken rule in 16 of 18 flawed networks (95% CI 67% to 97%), including all 15 with a confound or a threshold shift. It repaired each of the 16 by changing a median of 3 parameters. Fine-tuning on protocol-labeled requests also repaired every network, but changed nearly every parameter. Each confound and shift it found was visible in single units, and a simple statistic chosen for each flaw type found the same flaws. It cleared 2 of 5 healthy networks without being forced and found 1 of 3 triggers. Testing against the protocol detected every flaw, and black-box testing found the confounds and the shift faster. The results indicate that, in small networks, weight-space programming can locate and repair flaws that silence an input or shift a threshold, but cannot yet certify that a network is correct.
MedPhysBench: A Benchmark for Evaluating Language Models on Medical Physics Calculations and Error Detection
Omar Nusrat
AI 速览：作者构建了一个覆盖放疗、诊断影像、核医学与辐射防护的医学物理计算与查错基准，并报告了12个语言模型在留出测试集上的表现。
Medical physicists turn a prescription into a delivered dose and check one another's calculations, because one error in a beam calibration or a monitor-unit (MU) calculation can affect every patient treated with it. MedPhysBench has ten seeded generators of calculation items in radiation therapy, diagnostic imaging, nuclear medicine and radiation protection, up to clinical tasks such as a TG-51 calibration worksheet. Reference answers come from code implementing the governing protocols, and a grader converts units and applies tolerance bands. From the same inputs, separate checking items test error detection, showing a colleague's worksheet or only the colleague's result, half correct and half with one planted error. On a held-out test split of 1,200 items (half for two models), 12 models scored 33.3 to 100.0% on calculation. Of the nine answers graded incorrect for the four best (99.3-100.0%), five contained model errors and four followed NCRP Report 151 on points our items misstate or leave open. Below them, most models did worst on level-3 clinical tasks, and Claude Sonnet 5.5 omitted the inverse-square correction in 19 of 20 isocentric MU items. Five models detected at least 96.7% of mostly large planted errors with at most 4% false alarms; for four of the other seven, checking only the result raised false alarms significantly. In an exploratory targeted test with 20 public cases of an omitted inverse-square correction, the three best models tested caught it in every worksheet, and three weaker models accepted 16 to 18 of 20 planted worksheets, one of them while applying the correction in most of its own calculations. Claude Opus 5.5 accepted 6 of 20 wrong results, although it applied the correction in all 20 of its own. We report the effect of two problems in our shielding items, found after the runs, and release the code, per-item outcomes and targeted-test responses.
This work presents a new method for gradient-based optimization of the rhombic transposition flap and the excision it closes, which set the strain of the skin after closure. Previous studies swept at most three parameters. Here, using structural optimization methods, up to eight variables, two flap angles, two cut lengths and the size, shape and position of the rhombic excision, are optimized under up to 32 geometric and stress constraints. The skin is a pre-stretched, incompressible hyperelastic membrane that wrinkles instead of carrying compression. Smoothed level-sets on a fixed grid represent the cuts, and springs between interpolated points the sutures, so designs change without remeshing and the adjoint method gives the sensitivities. The energy the surgery adds to the skin is minimized, and the model reproduces the energies of a published 48-design sweep with a median deviation of 5%. The optimal flap angles, $(-23^\circ, 31^\circ)$, lie outside the swept range and add 63% less energy than the Limberg flap and 15% less than the best swept design. Under unequal in-vivo tension, the optimal flap turns with the direction of the tension. Fitting the excision to the lesion and its margin lowers the energy by a further 35%, and with all eight variables it is 78% below the Limberg flap. This optimum is the same for finer grids, narrower cuts and wider undermining, although wider undermining reduces its gain. Lowering its suture-line stress by 41% costs only 2.3% more energy.
Medical imaging has served as primary proving ground for clinical AI, yet a decade of intense research has not translated into proportionate bedside impact. We argue that this gap is not primarily a product of insufficient algorithmic performance, inadequate regulation, or limited explainability. Rather, it reflects a structural misalignment, between how AI systems are designed and evaluated, and how clinical decisions are made. This work identifies 8 interconnected dimensions of this misalignment: 1) the dominance of pixel-only models in a multimodal clinical world; 2) the erosion of physician trust through opaque and inflexible systems; 3) the unfulfilled promise of foundation models in data-sparse medical domains; 4) the persistent bottleneck of non-shareable under-curated datasets; 5) the gap between validated algorithms and deployable clinical platforms; 6) the slow, unmonitored translation of evidence into routine practice; 7) the risk of AI displacing rather than extending physician judgment; and 8) the failure of prediction-centric AI to generate actionable, agentic clinical guidance. For each dimension, we reframe the problem and propose a path forward, culminating in a vision of physician-centered AI solutions with agentic reasoning that extend, rather than replace, clinical judgment.
Mustafa Kadhim, Viktor Rogowski, Emilia Persson, Camila Gonzalez, André Haraldsson, Sofie Ceberg, Mikael Nilsson, Malin Kügele 等 10 位作者
Artificial intelligence is increasingly integrated into radiotherapy workflows, yet it remains vulnerable to out-of-distribution image data that can introduce unexpected behavior in clinical tasks. Deep learning (DL)-based anomaly detection for pelvic magnetic resonance imaging (MRI) remains limited and transparent evaluation for full automation is lacking. This study aimed to develop two self-supervised anomaly-detection framework variations in radiotherapy for pelvic and brain MRI.A two-stage self-supervised DL-framework was trained separately for pelvic and brain MRI utilizing three public datasets. The first stage compressed normal image data into discrete representations, and a second stage modeled the representation distribution. Anomaly-evidence was obtained by combining perceptual image differences with token-surprisal scores via negative log-likelihood estimation. Automated detection was evaluated for two anatomical cohorts: pelvis, encompassing both synthetic global anomalies and real clinical anomalies, and brain, comprising clinically annotated anomalies. Performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). For each cohort, a subset of normal images was reserved for analysis. Promising exploratory performance across evaluation cohorts was obtained. AUCs were 0.97 (95% CI 0.96-0.99) for synthetic pelvic anomalies, 0.95 (95% CI 0.90-1.00) for clinical pelvic anomalies, and 0.81 (95% CI 0.74-0.87) for brain anomalies. Qualitative heatmap analysis showed spatial agreement between detected anomalies and reference annotations. The framework variations showed promising anomaly detection performance across pelvic and brain MRI, supporting its potential for automated MRI quality control in radiotherapy. Heatmap-based visualization further improved interpretability by spatially localizing detected anomalies.
Computed tomography (CT) provides diverse patient information and various clinical tasks, for which numerous deep learning models have been investigated. Recent years have seen the development of vision-language foundation models that can comprehensively extract clinical features, driven by representation learning that combines CT volumes with radiology reports. Existing approaches typically adopt Contrastive Language-Image Pre-training (CLIP)-based representation learning, where pairs from different cases are, in principle, treated as negative pairs such that the volume and report embeddings are pushed apart. In the biomedical imaging, however, treating different cases as negative pairs is not always appropriate, particularly when clinical similarity exists. In this study, we propose a supervised representation learning framework grounded in clinical knowledge from language foundation models, rather than the weak supervision used in CLIP. Specifically, we utilized a large language model to extract abnormal finding phrases from reports. We then explicitly constructed the target features that the image encoder should output, as the semantic representation of the extracted findings using a biomedical text embedding model. Our Findings-aware Language-Image Pre-training (FiLIP) outperformed weakly supervised approaches in multi-label zero-shot finding classification in both in-domain and external-domain evaluations, and also proved effective when fine-tuned for the downstream task of radiology report generation.
Robust segmentation of biomedical images remains challenging primarily due to ill-defined object boundaries. Although several loss functions have recently been proposed to enhance object boundary recognition, concerns remain about the susceptibility to unstable annotations resulting from boundary ambiguities. Here, we introduce the Peripheral Prototypical Contrastive (PPC) loss, which leverages prototypical contrastive learning to improve boundary recognition and enable stable training even under noisy annotation conditions. The PPC loss utilizes prototype features obtained from local regions near object boundaries, focusing specifically on ambiguous areas that have been challenging for existing approaches. Extensive experiments across multiple datasets demonstrated that the PPC loss consistently improve segmentation performance even under conditions with noisy labels. A newly proposed strategy, which adaptively determines the local region for prototype calculation, enhanced robustness and outperformed existing methods. Our findings indicate the potential of the PPC loss in facilitating scalable and reliable medical image segmentation, including the development of foundation models with imperfect annotations.
Mike D Rinderknecht, Bernhard Fehlmann, Dimitar Stanev, Cedric Simillion, Ernst Bos, Letizia Leocani, Agne Kazlauskaite, Gary Cutter 等 10 位作者
Background: Digital health technologies allow for frequent, remote gait monitoring in people with multiple sclerosis (MS). However, to differentiate daily variability from actual disease progression in longitudinal data, established minimal clinically important differences (MCID) are required. Currently, there is limited literature defining these thresholds for digital gait metrics. Objective: To establish MCIDs for digital gait measures reflecting progression in MS. Methods: Digital gait measures were captured via daily, remote, smartphone-based Two-Minute Walk Tests in CONSONANCE (NCT03523858), a phase 3b study of ocrelizumab in progressive MS. Using an anchor-based approach, median changes from baseline at Week 96 on digital gait measures were computed for patients showing clinically meaningful worsening on either Timed 25-Foot Walk, Ambulation Score, Expanded Disability Status Scale, or 12-item Multiple Sclerosis Walking Scale. These changes were subsequently triangulated to derive the MCID estimates. Results: 243 patients with progressive MS (female: n=125 (51%); mean [SD] age: 49.3 [9.3]; mean [SD] EDSS: 4.8 [1.4]) had digital gait data available at baseline and Week 96. Median changes were generally consistent across anchors. Triangulated MCIDs are: Step Velocity = -0.16 m/s, Step Velocity Scaled to Walking Time = -0.18 m/s, Step Duration = 0.06 s, Step Length = -0.07 m, Total Number of Steps = -28, and Total Distance Walked = -24 m. Conclusion: These MCIDs provide a framework for interpreting meaningful gait changes and integrating digital measures into MS outcome evaluation. Beyond facilitating novel clinical trial endpoints to evaluate treatment efficacy, they enable objective, real-world monitoring to advance personalized patient care.
Measurement of small magnetic fields requires shielding. For example, magnetoencephalography (MEG) measures small fields arising within the brain, in the presence of environmental fields several orders of magnitude larger. A common solution is passive shielding, but its cost, bulk and limited effectiveness motivate the alternative approach of active shielding, which comes in two flavours: feed-back and feed-forward. In feed-back shielding (the most common design), coils are driven from sensors that measure the corrected field. In feed-forward shielding (less common), coils are driven from sensors that measure the uncorrected field. The latter has the potential of providing better performance, but its implementation is hobbled by an inevitable coupling between coils and reference sensors. This paper describes an approach that employs a data-driven ``decoupling matrix" to suppress the coupling, enabling greater flexibility to achieve effective shielding across a wide frequency range. The method is evaluated by simulation, first using a simplified "flat world" in which the principle is clearly visualised, and then based on the more realistic geometry of a standard MEG system. Performance appears to be limited only by spatial sampling and numerical resolution. The method, which is data-driven and does not require geometric calculations, tight-tolerance fabrication, or manual calibration, is complementary with standard feed-back shielding methods. Active shielding is of particular interest in the context of a new generation of optically-pumped magnetometers (OPM), as the lower cost and reduced constraints of these new systems relative to SQUID-based systems provide an incentive to likewise reduce the cost and constraints of shielding.
Mojtaba Safari, Parag J. Parikh, Shansong Wang, Mingzhe Hu, Philip Connell, Stanley L. Liauw, Ralph R. Weichselbaum, Zhen Tian 等 9 位作者
Background and purpose: We propose PGPI4MR (prior-guided, physics-informed 4D MRI reconstruction), a deep learning method that leverages information from other respiratory phases or a setup scan, the first 30 s of the same scan, to improve 4D MRI reconstruction without extra scan time. Materials and methods: PGPI4MR is an unrolled network that reconstructs 16 respiratory phases from 60, 30, or 15 s of data, and single 3D volumes, each from 0.29-0.49 s of data, using a setup-scan image. The networks were trained once, on other volunteers, and were not retrained for new volunteers. We evaluated PGPI4MR on measured k-space from three datasets: free-breathing MR-linac (1.5T) and abdominal (3T) scans, and thigh scans with cuff compression (1.5T). All methods, including comparative models, received identical samples and were scored similarly. Results: In vivo, PGPI4MR had the lowest held-out error for 4D MRI at every scan time, and its 16 phases were ready a median of 7.5-13 s after the scan on the 2 GPUs tested. Coupling the 16 respiratory phases reduced the error at 60 s by 9.1% compared with reconstructing each phase alone. For sub-second MRI, the error at 0.49 s was 15.5% lower than that of the best comparison method, classical reconstruction toward the same prior, and on the faster GPU, every 0.29 s window was acquired and reconstructed within 0.5 s. Without retraining, PGPI4MR matched the best methods on the 3T abdominal scans but not on the thigh scans; after retraining on thigh data, it reduced the held-out error by 6.6% relative to the best comparison method. Conclusion: Prior information allowed PGPI4MR to reconstruct 4D and sub-second MRI more accurately than compressed sensing and a published 4D network, within seconds and without extra scan time. This could support 4D MRI for treatment planning and 3D motion monitoring during MR-guided radiotherapy.
T1-T2 correlation MRI has strong potential to quantify microstructural compartments without imposing a biological model. Traditionally, full T1-T2 spectra per voxel are reconstructed with an ill-conditioned and time-consuming Inverse Laplace Transform. We propose to view this challenge as an object detection task: A Detection Transformer (DETR) takes 64 TI/TE signals and directly predicts a set of compartments specified by T1, T2, and signal fraction. We achieve a mean average precision mAP@10 of 76.8% on simulated data, outperforming both our previous model design (53.8%) and regularized NNLS (66.1%). We show that our DETR approach outperforms NNLS in almost every setting across compartment count, compartmental signal fraction, and SNR. Reconstruction is orders of magnitude faster than NNLS. The results are fully reproducible from our public code: github.com/SebastianEndtTHI/t1t2-detr
Ultra-low-field (ULF) Magnetic Resonance Imaging (MRI) offers a portable, low-cost alternative to conventional scanners by operating at field strengths below 10 mT. At these fields, two sensor technologies compete: quantum SQUID (Superconducting Quantum Interference Device) sensors, which provide high sensitivity but require cryogenic cooling, and classical induction coils, which are inexpensive and room-temperature but produce weaker signals. With conventional physics-based imaging, both suffer from low signal-to-noise ratios (SNR). Data-driven reconstruction has overcome similar limits in related inverse problems, but its use for ULF-MRI has been hindered by a lack of open datasets, benchmarks, and flexible simulators. To close this gap, we present OpenULF, a modular, open-source simulation toolkit that generates paired training data and supports end-to-end machine-learning (ML) reconstruction. OpenULF integrates a Bloch-equation-validated physics simulator, calibrated SQUID and coil receiver models, reconstruction ML networks, and an automated design-space exploration engine. Architecturally, a receiver abstract base class, a typed HDF5 contract that decouples physics from learning, and a resumable sweep engine make sensors, models, and datasets independently substitutable. Using OpenULF, we train 774 models across four anatomical datasets and conduct three design-space explorations. The effectiveness of ML reconstruction depends strongly on the dataset and the sensor. We identify a SQUID-to-coil crossover near 450 $μ$T and demonstrate that sensor hardware and ML model capacity can be traded off during system design. Sensor and reconstruction choices must therefore be co-designed: the best sensor is set by the dataset and field strength, not by SNR.
Cone beam computed tomography (CBCT) is widely used for image guided and adaptive radiotherapy, but scatter, noise, reconstruction artifacts, and inaccurate Hounsfield units limit its quantitative use. Recent diffusion approaches for CBCT to CT synthesis remain largely image driven, with limited use of the underlying acquisition physics. We propose EPC-RawDC, a conditional 3D latent diffusion framework that combines volumetric generative modeling with two complementary constraints derived from the same scanner specific CBCT forward model. First, geometry equivariant projection constraint exploits the rotational equivariance of the CBCT forward operator by supervising the synthesized CT through correspondingly transformed reference projections. Second, raw data consistency uses noisy projection measurements of the individual acquisition as scan specific supervision during training. The framework was evaluated on two independent clinical cohorts with distinct scanner geometries and projection sources. For JUST, the physics constraints used measured clinical raw CBCT projections acquired with an Elekta system, whereas for SynthRAD2025 the projection data were simulated from the reference CT using a Varian geometry with Poisson photon statistics. EPC-RawDC achieved an SSIM of 0.97 in both cohorts and improved PSNR over DDPM by 2.66 dB on SynthRAD2025 and 2.10 dB on JUST, while reducing mean MAE by 20.3% and 19.0%, respectively. On SynthRAD2025, projection consistency and equivariant errors were reduced by 55.0% and 41.4% relative to DDPM. These results demonstrate that combining a volumetric diffusion prior with scanner geometry and scan specific projection measurements improves quantitative CBCT to CT synthesis across both simulated and measured projection domains. Code available after publication at https://github.com/ALZAHRAALTALIB/EPC-RawDC.
Yuzhen Ding, Hongying Feng, Jianfeng Zhao, Jason Holmes, Martin Bues, Lisa A. McGee, Jean-Claude M. Rwigema, Nathan Y. Yu 等 17 位作者
Background and purpose: Intensity-modulated proton therapy (IMPT) enables conformal dose delivery for head and neck (H&N) cancers. However, its sensitivity to inter-fractional anatomical changes motivates online adaptations, where rapid and accurate Monte Carlo (MC) dose calculation remains a major bottleneck. Low-statistics MC simulations reduce computation time but introduce noise. This study aimed to develop a diffusion-transformer framework to recover high-statistics dose distributions from noisy MC simulations. Materials and methods: A diffusion-transformer-based dose denoising framework was developed. Eighty H&N patients were used for training and additional 10 H&N patients and 30 patients with lung, breast, and prostate cancers (10 per site) were used for evaluation. Noisy and high-statistics dose maps were generated using MCsquare. The model was trained using noisy dose maps and CT images as inputs and high-statistics dose maps as ground truth, with mean squared error, residual, regional mean absolute error (MAE), and dose-volume histogram (DVH) losses. Performance was evaluated using MAE, 3D gamma passing rates, and DVH indices. Results: Mean whole-body MAE ranged from 0.08 +/- 0.06 to 0.20 +/- 0.11 Gy (Relative Biological Effectiveness, RBE) (RBE = 1.1) across test cohorts. Mean whole-body 3 %/2 mm gamma passing rates ranged from 95.8 +/- 0.8 % to 97.6 +/- 0.9 %. DVH indices from high-statistics and predicted dose distributions showed close agreement for clinical target volumes and organs of interest. Conclusions: The diffusion-transformer-based framework denoised low-statistics MC dose distributions. Despite being trained on H&N data, the model generalized to lung, breast, and prostate cohorts without retraining.