In an interdisciplinary project involving machine learning experts, drone and helicopter pilots, WindINR is developed as a framework for continuous high- resolution local wind query and sparse-observation correction. The method combines specialized neural networks, surrogate high-resolution truth from computational fluid dynamics (CFD) simulations, and a method for assimilation of in-situ wind observations.
Operational weather products are usually delivered as gridded analyses and forecasts, but many decisions in complex terrain require only a small number of fast, accurate wind estimates at user-specified locations. During a helicopter approach, for example, the relevant question is often not the full wind field over a large domain, but the local gusts and potential turbulence along an approach corridor and around a landing zone. This provides support for situational awareness.
These local winds can vary strongly because of terrain- induced channeling, ridge acceleration, land–sea contrast, and boundary-layer structure. Sparse new observations from drones/UAVs or stations may also arrive immediately before the decision. The desired operating point is therefore not simply another dense high-resolution forecast, but a local wind-state estimator that can be queried continuously and corrected quickly.
Our new innovation WindINR provides continuous high-resolution local wind queries with the ability for sparse-observation correction. The method combines implicit neural representation networks, physical numerical simulations, and in-situ assimilation of in-situ wind observations. The aim is for the in-situ observations to be drone/UAV acquired within a context of aviation, e.g. for helicopter search and rescue.
The project is a collaboration between UiT The Arctic University of Norway, EPFL, MBZUAI, and Tsinghua University.
Read more here: http://arxiv.org/abs/2605.09511
[18F]FDG Positron Emission Tomography (PET) is a widely used medical imaging technique that provides highly sensitive spatial representations of radiopharmaceuticals metabolized by tissues and cells. PET imaging is particularly valuable for identifying active tissues and detecting malignant growths, such as cancerous tumors, which appear as bright spots in the images. It is also instrumental in monitoring cancer therapies.
Traditionally, PET images are static, constructed by averaging gamma-ray counts over time after the radiotracer uptake stabilizes post-injection.While static PET is effective for many applications, its limitations become apparent in tasks like drug development, tracer design, and disease therapy research. Dynamic PET (dPET) offers an alternative by capturing radiation detections over the entire imaging duration and averaging them into smaller temporal bins. This results in a spatial time-series representation of tracer uptake, visualized through time-activity curves (TACs). These TACs reflect the temporal variability of tracer binding to proteins, free states in plasma, or metabolism. Tracer kinetic modeling, often using compartment models, estimates uptake rates between plasma, tissues, and cells. This process relies on the arterial input function (AIF), which represents plasma blood concentration.
However, obtaining an accurate AIF is challenging. The gold standard involves invasive arterial cannulation, which is complex, time-intensive, and unsuitable for long-term studies, especially in mice, as it requires terminal procedures like carotid artery sampling.
To address these challenges, non-invasive AIF estimation methods have been explored. Techniques such as population-based input functions and image-derived input functions (IDIFs) use TACs from delineated arteries or blood pools (e.g., the left ventricle). However, these methods require rigorous calibration to mitigate partial volume effects and motion artifacts, limiting their practicality. Recent advancements include machine learning-based input functions, but these approaches often depend on manual TAC delineations.In this project, we develop deep learning-based input function (DLIF) estimation methods for AIF.
Deep learning models can process entire dPET datasets and learn mappings to mimic IDIFs, including necessary calibrations. Additionally, we are advancing physics-informed deep learning models that integrate the physical dynamics of tracer uptake. These models enable voxel-wise kinetic modeling, eliminating the need for 3D segmentation steps. By combining deep learning with compartment models, we aim to provide a robust mathematical framework for describing tracer kinetics, offering insights into cellular and tissue activity at the voxel or region-of-interest level.
Read more here: https://www.frontiersin.org/journals/nuclear-medicine/articles/10.3389/fnume.2024.1372379/full
