Search for: All records

Natural images depict real-world scenes such as landscapes, animals, and everyday items. Transformer-based detectors, such as the Detection Transformer, have demonstrated strong object detection performance on natural image datasets. These models are typically optimized through complex engineering strategies tailored to the characteristics of natural scenes. However, medical imaging presents unique challenges, such as high resolutions, smaller and fewer regions of interest, and subtle inter-class differences, which differ significantly from natural images. In this study, we evaluated the effectiveness of common design choices in transformer-based detectors when applied to medical imaging. Using two representative datasets, a mammography dataset and a chest CT dataset, we showed that common design choices proposed for natural images, including complex encoder architectures, multi-scale feature fusion, query initialization, and iterative bounding box refinement, fail to improve and can even be detrimental to the object detection performance. In contrast, simpler and shallower architectures often achieve equal or superior results with less computational cost. These findings highlight that standard design practices need to be reconsidered when adapting transformer models to medical imaging, and suggest that simplicity may be more effective than added complexity in this domain. Our model code and weights are publicly available at https://github.com/nyukat/Mammo-DETR 

The Segment Anything Model (SAM) is a large-scale foundation model that has revolutionized segmentation methodology. Despite its impressive generalization ability, the segmentation accuracy of SAM on images with intricate structures is often unsatisfactory. Recent works have proposed lightweight fine-tuning using high-quality annotated data to improve accuracy on such images. However, here we provide extensive empirical evidence that this strategy leads to forgetting how to "segment anything": these models lose the original generalization abilities of SAM, in the sense that they perform worse for segmentation tasks not represented in the annotated fine-tuning set. To improve performance without forgetting, we introduce a novel framework that combines high-quality annotated data with a large unlabeled dataset. The framework relies on two methodological innovations. First, we quantify the uncertainty in the SAM pseudo labels associated with the unlabeled data and leverage it to perform uncertainty-aware fine-tuning. Second, we encode the type of segmentation task associated with each training example using a task prompt to reduce ambiguity. We evaluated the proposed Segmentation with Uncertainty Model (SUM) on a diverse test set consisting of 14 public benchmarks, where it achieves state-of-the-art results. Notably, our method consistently surpasses SAM by 3-6 points in mean IoU and 4-7 in mean boundary IoU across point-prompt interactive segmentation rounds. 

Ocean mesoscale eddies are often poorly represented in climate models, and therefore, their effects on the large scale circulation must be parameterized. Traditional parameterizations, which represent the bulk effect of the unresolved eddies, can be improved with new subgrid models learned directly from data. Zanna and Bolton (ZB20) applied an equation‐discovery algorithm to reveal an interpretable expression parameterizing the subgrid momentum fluxes by mesoscale eddies through the components of the velocity‐gradient tensor. In this work, we implement the ZB20 parameterization into the primitive‐equation GFDL MOM6 ocean model and test it in two idealized configurations with significantly different dynamical regimes and topography. The original parameterization was found to generate excessive numerical noise near the grid scale. We propose two filtering approaches to avoid the numerical issues and additionally enhance the strength of large‐scale energy backscatter. The filtered ZB20 parameterizations led to improved climatological mean state and energy distributions, compared to the current state‐of‐the‐art energy backscatter parameterizations. The filtered ZB20 parameterizations are scale‐aware and, consequently, can be used with a single value of the non‐dimensional scaling coefficient for a range of resolutions. The successful application of the filtered ZB20 parameterizations to parameterize mesoscale eddies in two idealized configurations offers a promising opportunity to reduce long‐standing biases in global ocean simulations in future studies. 

Materials functionalities may be associated with atomic-level structural dynamics occurring on the millisecond timescale. However, the capability of electron microscopy to image structures with high spatial resolution and millisecond temporal resolution is often limited by poor signal-to-noise ratios. With an unsupervised deep denoising framework, we observed metal nanoparticle surfaces (platinum nanoparticles on cerium oxide) in a gas environment with time resolutions down to 10 milliseconds at a moderate electron dose. On this timescale, many nanoparticle surfaces continuously transition between ordered and disordered configurations. Stress fields can penetrate below the surface, leading to defect formation and destabilization, thus making the nanoparticle fluxional. Combining this unsupervised denoiser with in situ electron microscopy greatly improves spatiotemporal characterization, opening a new window for the exploration of atomic-level structural dynamics in materials. 

Subgrid parameterizations of mesoscale eddies continue to be in demand for climate simulations. These subgrid parameterizations can be powerfully designed using physics and/or data‐driven methods, with uncertainty quantification. For example, Guillaumin and Zanna (2021) proposed a Machine Learning (ML) model that predicts subgrid forcing and its local uncertainty. The major assumption and potential drawback of this model is the statistical independence of stochastic residuals between grid points. Here, we aim to improve the simulation of stochastic forcing with generative models of ML, such as Generative adversarial network (GAN) and Variational autoencoder (VAE). Generative models learn the distribution of subgrid forcing conditioned on the resolved flow directly from data and they can produce new samples from this distribution. Generative models can potentially capture not only the spatial correlation but any statistically significant property of subgrid forcing. We test the proposed stochastic parameterizations offline and online in an idealized ocean model. We show that generative models are able to predict subgrid forcing and its uncertainty with spatially correlated stochastic forcing. Online simulations for a range of resolutions demonstrated that generative models are superior to the baseline ML model at the coarsest resolution.