Search

Purpose: Beam angle selection is critical in proton therapy treatment planning, yet automated approaches remain underexplored. This study presents and evaluates GAMBAS, a novel, fast machine learning model for automatic beam angle selection. Methods: The model extracts a predefined set of anatomical features from a patient's CT and structure contours. Using these features, it identifies the most similar patient from a training database and suggests that patient's beam arrangement. A retrospective study with 19 patients was conducted, comparing this model's suggestions to human planners' choices and randomly selected beam arrangements from the training dataset. An expert treatment planner evaluated the plans on quality (scale 1-5), ranked them, and guessed the method used. Results: The number of acceptable (score 4 or 5) plans was comparable between human-chosen 17 (89%) and model-selected 16(84%) beam arrangements. The fully automatic treatment planning took between 4 - 7 min (mean 5 min). Conclusion: The model produces beam arrangements of comparable quality to those chosen by human planners, demonstrating its potential as a fast tool for quality assurance and patient selection, although it is not yet ready for clinical use., Comment: 8 pages, 2 figures, 1 table

The importance of computers is continually increasing in radiotherapy. Efficient algorithms, implementations and the ability to leverage advancements in computer science are crucial to improve cancer care even further and deliver the best treatment to each patient. Yet, the software landscape for radiotherapy is fragmented into proprietary systems that do not share a common interface. Further, the radiotherapy community does not have access to the vast possibilities offered by modern programming languages and their ecosystem of libraries yet. We present JulianA.jl, a novel Julia package for radiotherapy. It aims to provide a modular and flexible foundation for the development and efficient implementation of algorithms and workflows for radiotherapy researchers and clinicians. JulianA.jl can be interfaced with any scriptable treatment planning system, be it commercial, open source or in-house developed. This article highlights our design choices and showcases the package's simplicity and powerful automatic treatment planning capabilities.

Background: Automatic treatment planning promises many benefits for both research and clinical environments. For clinics, autoplanning promises to reduce planning time and achieve more comparable treatment plans and thereby reduce inter-planner variability. Further, it can assist clinicians in quality assurance by providing a minimum plan quality standard. Finally, autoplanning is an essential part of patient selection, which is crucial for the advancement of proton therapy itself. Methods: A retrospective planning study using a cohort of 17 head-and-neck cancer patients treated at our institute. The clinically accepted plans created by dosimetrists (d-plans) were compared to automatically generated JulianA plans (j-plans). Both methods used the same beam arrangement. The plans were analysed by two expert reviewers without knowing how each plan was created. They assessed the plan quality and stated a preference. Results: All of the j-plans were deemed rather or clearly acceptable, resulting in a higher acceptability than the d-plans. The j-plan was considered superior in 14 (82.4%) cases, of equal quality for 1 (5.9%) and inferior to the d-plan for only 2 (11.8%) of the cases. The reviewers concluded that JulianA achieves more conformal dose distributions for the 15 (88.2%) cases where the j-plans were at least as good as the d-plans. Conclusions: The results show that the JulianA is ready to be used as a clinical quality assurance tool and research platform at our institute. While these results are encouraging, further research is needed to reduce the number of spots further and introduce robustness considerations into the optimisation algorithm in order to employ it on a daily basis for patient treatment.

Motion information from 4D medical imaging offers critical insights into dynamic changes in patient anatomy for clinical assessments and radiotherapy planning and, thereby, enhances the capabilities of 3D image analysis. However, inherent physical and technical constraints of imaging hardware often necessitate a compromise between temporal resolution and image quality. Frame interpolation emerges as a pivotal solution to this challenge. Previous methods often suffer from discretion when they estimate the intermediate motion and execute the forward warping. In this study, we draw inspiration from fluid mechanics to propose a novel approach for continuously modeling patient anatomic motion using implicit neural representation. It ensures both spatial and temporal continuity, effectively bridging Eulerian and Lagrangian specifications together to naturally facilitate continuous frame interpolation. Our experiments across multiple datasets underscore the method's superior accuracy and speed. Furthermore, as a case-specific optimization (training-free) approach, it circumvents the need for extensive datasets and addresses model generalization issues.

Background and purpose: Deformable image registration (DIR) is a crucial tool in radiotherapy for extracting and modelling organ motion. However, when significant changes and sliding boundaries are present, it faces compromised accuracy and uncertainty, determining the subsequential contour propagation and dose accumulation procedures. Materials and methods: We propose an implicit neural representation (INR)-based approach modelling motion continuously in both space and time, named Continues-sPatial-Temporal DIR (CPT-DIR). This method uses a multilayer perception (MLP) network to map 3D coordinate (x,y,z) to its corresponding velocity vector (vx,vy,vz). The displacement vectors (dx,dy,dz) are then calculated by integrating velocity vectors over time. The MLP's parameters can rapidly adapt to new cases without pre-training, enhancing optimisation. The DIR's performance was tested on the DIR-Lab dataset of 10 lung 4DCT cases, using metrics of landmark accuracy (TRE), contour conformity (Dice) and image similarity (MAE). Results: The proposed CPT-DIR can reduce landmark TRE from 2.79mm to 0.99mm, outperforming B-splines' results for all cases. The MAE of the whole-body region improves from 35.46HU to 28.99HU. Furthermore, CPT-DIR surpasses B-splines for accuracy in the sliding boundary region, lowering MAE and increasing Dice coefficients for the ribcage from 65.65HU and 90.41% to 42.04HU and 90.56%, versus 75.40HU and 89.30% without registration. Meanwhile, CPT-DIR offers significant speed advantages, completing in under 15 seconds compared to a few minutes with the conventional B-splines method. Conclusion: Leveraging the continuous representations, the CPT-DIR method significantly enhances registration accuracy, automation and speed, outperforming traditional B-splines in landmark and contour precision, particularly in the challenging areas.

In recent advancements in proton therapy, MR-based treatment planning is gaining momentum to minimize additional radiation exposure compared to traditional CT-based methods. This transition highlights the critical need for accurate MR-to-CT image synthesis, which is essential for precise proton dose calculations. Our research introduces the Diffusion Schr\"odinger Bridge Models (DSBM), an innovative approach for high-quality MR-to-CT synthesis. DSBM learns the nonlinear diffusion processes between MR and CT data distributions. This method improves upon traditional diffusion models by initiating synthesis from the prior distribution rather than the Gaussian distribution, enhancing both generation quality and efficiency. We validated the effectiveness of DSBM on a head and neck cancer dataset, demonstrating its superiority over traditional image synthesis methods through both image-level and dosimetric-level evaluations. The effectiveness of DSBM in MR-based proton treatment planning highlights its potential as a valuable tool in various clinical scenarios., Comment: International Conference on the use of Computers in Radiation therapy (ICCR)

Most earthquake energy release arises during fault slip many kilometers below the Earth's surface. Understanding earthquakes and their hazard requires mapping the geometry and distribution of this slip. Such finite-fault maps are typically derived from surface phenomena, such as seismic and geodetic ground motions. Here we introduce an imaging procedure for mapping finite-fault slip directly from seismicity and aftershocks - phenomena occurring at depth around an earthquake rupture. For specified source and receiver faults, we map source-fault slip in 3D by correlation of point-source Coulomb failure stress change (${\Delta}$CFS) kernels across the distribution of seismicity around an earthquake. These seismicity-stress maps show relative, static fault slip compatible with the surrounding seismicity given the physics of ${\Delta}$CFS; they can aid other slip inversions and aftershock forecasting, and be obtained for early instrumental earthquakes. We verify this procedure recovers synthetic fault slip, and matches independent estimates of slip for the 2004 Mw 6.0 Parkfield and 2021 Mw 6.0 Antelope Valley California earthquakes. For the 2018 Mw 7.1 Anchorage Alaska intra-slab earthquake, seismicity-stress maps, combined with multi-scale precise hypocenter relocation, resolve the enigma of which mainshock faulting plane ruptured (the gently east-dipping plane), and clarify slab structures activated in the energetic aftershock sequence., Comment: 41 pages, 12 figures; for supplementary movies, see https://zenodo.org/doi/10.5281/zenodo.10939981 ; presented at 2024 Annual Meeting, Seismological Society of America, 29 April-3 May, Anchorage, Alaska. v2: updated synthetic fault slip analysis, minor text changes

This appendix volume supplements "Evaluating the Federal Innovative Assessment Demonstration Authority: Early Implementation and Progress of State Efforts to Develop New Statewide Academic Assessments. Evaluation Report. NCEE 2023-004" analyzing the early progress of the first five assessment systems participating in the Innovative Assessment Demonstration Authority (IADA) created by the 2015 reauthorization of the Elementary and Secondary Education Act. These systems include: (1) the New Hampshire Performance Assessment of Competency Education (NH PACE); (2) the Louisiana Educational Assessment Program 2025 Humanities (LEAP 2025 Humanities); (3) the North Carolina Personalized Assessment Tool (NC PAT); (4) the Georgia MAP Assessment Partnership (GMAP) through-year assessment and (5) the Georgia Navvy assessment system. The volume documents the features of the IADA program and the five IADA systems, the study team's research activities, and includes more detailed findings than were summarized in the report. [For the full report, see ED627872. For the Study Highlights, see ED627880.]

Education officials have long hoped that the statewide academic assessments most students take each year could be used not only for accountability but also to guide instruction. Congress established the Innovative Assessment Demonstration Authority (IADA) program in 2015 to help address this goal, offering up to seven states temporary flexibility from federal testing requirements so that they may more easily make progress toward replacing their current assessments with more innovative ones. The key incentive to participate in IADA is that students trying out the innovative assessment are not required to also take the state's current assessment. However, states approved for IADA must still show that their innovative assessments meet most requirements for federal accountability, and they are expected to implement the new assessments statewide within 5 years. This report describes the progress of the first five assessment systems approved under IADA in order to help policymakers consider expanding the program to more states. The report is primarily based on an analysis of states' IADA applications and performance reports to the U.S. Department of Education through the 2020-2021 school year and is part of a broader evaluation of IADA required by Congress. [For the Appendix, see ED627873. For the Study Highlights, see ED627880.]

Intra-fraction motion in radiotherapy is commonly modeled using deformable image registration (DIR). However, existing methods often struggle to balance speed and accuracy, limiting their applicability in clinical scenarios. This study introduces a novel approach that harnesses Neural Graphics Primitives (NGP) to optimize the displacement vector field (DVF). Our method leverages learned primitives, processed as splats, and interpolates within space using a shallow neural network. Uniquely, it enables self-supervised optimization at an ultra-fast speed, negating the need for pre-training on extensive datasets and allowing seamless adaptation to new cases. We validated this approach on the 4D-CT lung dataset DIR-lab, achieving a target registration error (TRE) of 1.15\pm1.15 mm within a remarkable time of 1.77 seconds. Notably, our method also addresses the sliding boundary problem, a common challenge in conventional DIR methods.

Background: In MR-guided proton therapy planning, aligning MR and CT images is key for MR-based CT synthesis, especially in mobile regions like the head-and-neck. Misalignments here can lead to less accurate synthetic CT (sCT) images, impacting treatment precision. Purpose: This study introduces a novel network that cohesively unifies image generation and registration processes to enhance the quality and anatomical fidelity of sCTs derived from better-aligned MR images. Methods: The approach synergizes a generation network (G) with a deformable registration network (R), optimizing them jointly in MR-to-CT synthesis. This goal is achieved by alternately minimizing the discrepancies between the generated/registered CT images and their corresponding reference CT counterparts. The generation network employs a UNet architecture, while the registration network leverages an implicit neural representation of the Deformable Vector Fields (DVFs). We validated this method on a dataset comprising 60 Head-and-Neck patients, reserving 12 cases for holdout testing. Results: Compared to the baseline Pix2Pix method with MAE 124.95\pm 30.74 HU, the proposed technique demonstrated 80.98\pm 7.55 HU. The unified translation-registration network produced sharper and more anatomically congruent outputs, showing superior efficacy in converting MR images to sCTs. Additionally, from a dosimetric perspective, the plan recalculated on the resulting sCTs resulted in a remarkably reduced discrepancy to the reference proton plans. Conclusions: This study conclusively demonstrates that a holistic MR-based CT synthesis approach, integrating both image-to-image translation and deformable registration, significantly improves the precision and quality of sCT generation, particularly for the challenging body area with varied anatomic changes between corresponding MR and CT.

Creating high quality treatment plans is crucial for a successful radiotherapy treatment. However, it demands substantial effort and special training for dosimetrists. Existing automated treatment planning systems typically require either an explicit prioritization of planning objectives, human-assigned objective weights, large amounts of historic plans to train an artificial intelligence or long planning times. Many of the existing auto-planning tools are difficult to extend to new planning goals. A new spot weight optimisation algorithm, called JulianA, was developed. The algorithm minimises a scalar loss function that is built only based on the prescribed dose to the tumour and organs at risk (OARs), but does not rely on historic plans. The objective weights in the loss function have default values that do not need to be changed for the patients in our dataset. The system is a versatile tool for researchers and clinicians without specialised programming skills. Extending it is as easy as adding an additional term to the loss function. JulianA was validated on a dataset of 19 patients with intra- and extracerebral neoplasms within the cranial region that had been treated at our institute. For each patient, a reference plan which was delivered to the cancer patient, was exported from our treatment database. Then JulianA created the auto plan using the same beam arrangement. The reference and auto plans were given to a blinded independent reviewer who assessed the acceptability of each plan, ranked the plans and assigned the human-/machine-made labels. The auto plans were considered acceptable in 16 out of 19 patients and at least as good as the reference plan for 11 patients. Whether a plan was crafted by a dosimetrist or JulianA was only recognised for 9 cases. The median time for the spot weight optimisation is approx. 2 min (range: 0.5 min - 7 min).

We present light curves and flares from a seven day, multi-wavelength observational campaign of AU Mic, a young and active dM1e star with exoplanets and a debris disk. We report on 73 unique flares between the X-ray to optical data. We use high-time resolution NUV photometry and soft X-ray (SXR) data from XMM-Newton to study the empirical Neupert effect, which correlates the gradual and impulsive phase flaring emissions. We find that 65% (30 of 46) flares do not follow the Neupert effect, which is three times more excursions than seen in solar flares, and propose a four part Neupert effect classification (Neupert, Quasi-Neupert, Non-Neupert I & II) to explain the multi-wavelength responses. While the SXR emission generally lags behind the NUV as expected from the chromospheric evaporation flare models, the Neupert effect is more prevalent in larger, more impulsive flares. Preliminary flaring rate analysis with X-ray and U-band data suggests that previously estimated energy ratios hold for a collection of flares observed over the same time period, but not necessarily for an individual, multi-wavelength flare. These results imply that one model cannot explain all stellar flares and care should be taken when extrapolating between wavelength regimes. Future work will expand wavelength coverage using radio data to constrain the nonthermal empirical and theoretical Neupert effects to better refine models and bridge the gap between stellar and solar flare physics., Comment: 46 pages, 18 figures, 15 tables. Accepted for publication in The Astrophysical Journal

Objective: The aim of this work is to investigate the feasibility of the Jagiellonian Positron Emission Tomography (J-PET) scanner for intra-treatment proton beam range monitoring. Approach: The Monte Carlo simulation studies with GATE and PET image reconstruction with CASToR were performed in order to compare six J-PET scanner geometries (three dual-heads and three cylindrical). We simulated proton irradiation of a PMMA phantom with a Single Pencil Beam (SPB) and Spread-Out Bragg Peak (SOBP) of various ranges. The sensitivity and precision of each scanner were calculated, and considering the setup's cost-effectiveness, we indicated potentially optimal geometries for the J-PET scanner prototype dedicated to the proton beam range assessment. Main results: The investigations indicate that the double-layer cylindrical and triple-layer double-head configurations are the most promising for clinical application. We found that the scanner sensitivity is of the order of 10$^{-5}$ coincidences per primary proton, while the precision of the range assessment for both SPB and SOBP irradiation plans was found below 1 mm. Among the scanners with the same number of detector modules, the best results are found for the triple-layer dual-head geometry. Significance: We performed simulation studies demonstrating that the feasibility of the J-PET detector for PET-based proton beam therapy range monitoring is possible with reasonable sensitivity and precision enabling its pre-clinical tests in the clinical proton therapy environment. Considering the sensitivity, precision and cost-effectiveness, the double-layer cylindrical and triple-layer dual-head J-PET geometry configurations seem promising for the future clinical application. Experimental tests are needed to confirm these findings., Comment: 19 pages, 7 pages

Grayscale lithography allows the creation of micrometer-scale features with spatially-controlled height in a process that is fully compatible with standard lithography. Here, solid immersion lenses are demonstrated in silicon carbide using a novel fabrication protocol combining grayscale lithography and hard-mask techniques to allow nearly hemispherical lenses of 5 $\mu$m radius to be etched into the substrate. The technique is highly scalable and compatible with CMOS technology, and device aspect ratios can be tuned after resist patterning by controlling the chemistry of the subsequent dry etch. These results provide a low-cost, high-throughput and industrially-relevant alternative to focused ion beam milling for the creation of high-aspect-ratio, rounded microstructures., Comment: 9 pages, 4 + 2 supplementary figures

Purpose In PBS particle therapy, short treatment delivery time is paramount for the efficient treatment of moving targets with motion mitigation techniques (such as breath-hold, rescanning, and gating). Energy and spot position change time are limiting factors in reducing treatment time. In this study, we designed a universal and dynamic energy modulator (ridge filter, RF) to broaden the Bragg peak, to reduce the number of energies and spots required to cover the target volume, thus lowering the treatment time. Methods Our RF unit comprises two identical RFs placed just before the isocenter. Both RFs move relative to each other, changing the Bragg peaks characteristics dynamically. We simulated different Bragg peak shapes with the RF in TOPAS and validated them experimentally. We then delivered single-field plans with 1Gy/fraction to different geometrical targets in water, to measure the dose delivery time using the RF and compare it with the clinical settings. Results Aligning the RFs in different positions produces different broadening in the Bragg peak; we achieved a maximum broadening of 2 cm. With RF we reduced the number of energies in a field by more than 60%, and the dose delivery time by 50%, for all geometrical targets investigated, without compromising the dose distribution transverse and distal fall-off. Conclusions Our novel universal and dynamic RF allows for the adaptation of the Bragg peak broadening for a spot and/or energy layer based on the requirement of dose shaping in the target volume. It significantly reduces the number of energy layers and spots to cover the target volume, and thus the treatment time. This RF design is ideal for ultra-fast treatment delivery within a single breath-hold (5-10 sec), efficient delivery of motion mitigation techniques, and small animal irradiation with ultra-high dose rates (FLASH)., Comment: 18 pages, 9 figures

The winds of massive stars are important for their direct impact on the interstellar medium, and for their influence on the final state of a star prior to it exploding as a supernova. However, the dynamics of these winds is understood primarily via their illumination from a single central source. The Doppler shift seen in resonance lines is a useful tool for inferring these dynamics, but the mapping from that Doppler shift to the radial distance from the source is ambiguous. Binary systems can reduce this ambiguity by providing a second light source at a known radius in the wind, seen from orbitally modulated directions. From the nature of the collision between the winds, a massive companion also provides unique additional information about wind momentum fluxes. Since massive stars are strong ultraviolet (UV) sources, and UV resonance line opacity in the wind is strong, UV instruments with a high resolution spectroscopic capability are essential for extracting this dynamical information. Polarimetric capability also helps to further resolve ambiguities in aspects of the wind geometry that are not axisymmetric about the line of sight, because of its unique access to scattering direction information. We review how the proposed MIDEX-scale mission Polstar can use UV spectropolarimetric observations to critically constrain the physics of colliding winds, and hence radiatively-driven winds in general. We propose a sample of 20 binary targets, capitalizing on this unique combination of illumination by companion starlight, and collision with a companion wind, to probe wind attributes over a range in wind strengths. Of particular interest is the hypothesis that the radial distribution of the wind acceleration is altered significantly, when the radiative transfer within the winds becomes optically thick to resonance scattering in multiple overlapping UV lines., Comment: 26 pages, 12 figures, Review in a topical collection series of Astrophysics and Space Sciences on the proposed Polstar satellite. arXiv admin note: substantial text overlap with arXiv:2111.11552

This resource focuses on recent changes in the data sources and measurements of Part B Indicators 1 and 2. It offers a side-by-side comparison of the State Performance Plan/Annual Performance Report (SPP/APR) Part B Indicator measurement tables for FFY 2019 and FFY 2020-2025 to highlight relevant differences in SPP/APR reporting of graduation rate and dropout rate. The resource specifically addresses the treatment of "alternate diploma" in the new calculation. For FFY 2019, the calculation of graduation rate included students receiving an alternate diploma in the numerator. For FFY 2020-2025, the calculation of graduation rate includes students receiving an alternate diploma in the denominator. The calculation for Indicator 2 remains similar from FFY 2019 to FFY 2020-2025; however, it explicitly adds students receiving an alternate diploma in the denominator.

This resource describes new reporting requirements and measurement changes to Part B State Performance Plan and Annual Performance Report (SPP/APR) Indicator 3. The majority of students with disabilities participate in regular statewide assessments, and a small percentage of students with the most significant disabilities participate in alternate assessments. In FFY 2020, OSEP requires states to disaggregate statewide assessment results for these two groups of students with disabilities (i.e., those who participate in the regular assessment and those who participate in the alternate assessment). OSEP also requires states to calculate a "proficiency rate gap" between students with and without disabilities who participate in the regular statewide assessment. The resource offers a side-by-side comparison of the SPP/APR Part B indicator measurement tables for FFY 2019 and FFY 2020 to highlight relevant differences in SPP/APR reporting of Indicator 3.

Massive star binaries are critical laboratories for measuring masses and stellar wind mass-loss rates. A major challenge is inferring viewing inclination and extracting information about the colliding wind interaction (CWI) region. Polarimetric variability from electron scattering in the highly ionized winds provides important diagnostic information about system geometry. We combine for the first time the well-known generalized treatment of \citet{brown_polarisation_1978} for variable polarization from binaries with the semi-analytic solution for the geometry and surface density CWI shock interface between the winds based on Canto et al 1996. Our calculations include some simplifications in the form of inverse square-law wind densities and the assumption of axisymmetry, but in so doing arrive at several robust conclusions. One is that when the winds are nearly equal (e.g., O\,+\,O binaries), the polarization has a relatively mild decline with binary separation. Another is that despite Thomson scattering being a gray opacity, the continuum polarization can show chromatic effects at ultraviolet wavelengths but will be mostly constant at longer wavelengths. Finally, when one wind dominates the other, as for example in WR+OB binaries, the polarization is expected to be larger at wavelengths where the OB component is more luminous, and generally smaller at wavelengths where the WR component is more luminous. This behavior arises because from the perspective of the WR star, the distortion of the scattering envelope from spherical is a minor perturbation situated far from the WR star. By contrast, the polarization contribution from the OB star is dominated by the geometry of the CWI shock., Comment: accepted for publication in the Astrophysical Journal. arXiv admin note: text overlap with arXiv:2111.11552

The COVID-19 pandemic has forced school leaders to make difficult decisions about how to balance students' educational needs with the health, safety, and social-emotional well-being of students and teachers. The American Rescue Plan Act of 2021 (Rescue Plan) recently enacted by Congress offers unprecedented amounts of potentially transformative funding to elementary and secondary schools. The Rescue Plan has the potential to simultaneously spur post-pandemic learning recovery efforts and address long-standing socio-demographic gaps in academic achievement and opportunity. This brief is intended to help state and local education leaders consider how to approach this opportunity to meet a persistent problem: the need for more educators of color to serve students. To do this, the authors encourage local education agencies (LEAs) to act strategically by leveraging the findings and evidence-based practices (EBP) of two distinct research bases: (1) High-dosage tutoring, extended-day programs, and summer learning programs are effective options for learning recovery; and (2) Students benefit from exposure to a diverse, representative educator workforce.

Modern livestock farming is increasingly data driven and frequently relies on efficient remote sensing to gather data over wide areas. High resolution satellite imagery is one such data source, which is becoming more accessible for farmers as coverage increases and cost falls. Such images can be used to detect and track animals, monitor pasture changes, and understand land use. Many of the data driven models being applied to these tasks require ground truthing at resolutions higher than satellites can provide. Simultaneously, there is a lack of available aerial imagery focused on farmland changes that occur over days or weeks, such as herd movement. With this goal in mind, we present a new multi-temporal dataset of high resolution UAV imagery which is artificially degraded to match satellite data quality. An empirical blurring metric is used to calibrate the degradation process against actual satellite imagery of the area. UAV surveys were flown repeatedly over several weeks, for specific farm locations. This 5cm/pixel data is sufficiently high resolution to accurately ground truth cattle locations, and other factors such as grass cover. From 33 wide area UAV surveys, 1869 patches were extracted and artificially degraded using an accurate satellite optical model to simulate satellite data. Geographic patches from multiple time periods are aligned and presented as sets, providing a multi-temporal dataset that can be used for detecting changes on farms. The geo-referenced images and 27,853 manually annotated cattle labels are made publicly available., Comment: 16 pages, 7 figures, 2 tables

Synthetic datasets are often presented as a silver-bullet solution to the problem of privacy-preserving data publishing. However, for many applications, synthetic data has been shown to have limited utility when used to train predictive models. One promising potential application of these data is in the exploratory phase of the machine learning workflow, which involves understanding, engineering and selecting features. This phase often involves considerable time, and depends on the availability of data. There would be substantial value in synthetic data that permitted these steps to be carried out while, for example, data access was being negotiated, or with fewer information governance restrictions. This paper presents an empirical analysis of the agreement between the feature importance obtained from raw and from synthetic data, on a range of artificially generated and real-world datasets (where feature importance represents how useful each feature is when predicting a the outcome). We employ two differentially-private methods to produce synthetic data, and apply various utility measures to quantify the agreement in feature importance as this varies with the level of privacy. Our results indicate that synthetic data can sometimes preserve several representations of the ranking of feature importance in simple settings but their performance is not consistent and depends upon a number of factors. Particular caution should be exercised in more nuanced real-world settings, where synthetic data can lead to differences in ranked feature importance that could alter key modelling decisions. This work has important implications for developing synthetic versions of highly sensitive data sets in fields such as finance and healthcare., Comment: 27 pages, 8 figures