The rapid evolution of software libraries poses a considerable hurdle for code generation, necessitating continuous adaptation to frequent version updates while preserving backward compatibility. While existing code evolution benchmarks provide valuable insights, they typically lack execution-based evaluation for generating code compliant with specific library versions. To address this, we introduce GitChameleon 2.0, a novel, meticulously curated dataset comprising 328 Python code completion problems, each conditioned on specific library versions and accompanied by executable unit tests. GitChameleon 2.0 rigorously evaluates the capacity of contemporary large language models (LLMs), LLM-powered agents, code assistants, and RAG systems to perform version-conditioned code generation that demonstrates functional accuracy through execution. Our extensive evaluations indicate that state-of-the-art systems encounter significant challenges with this task; enterprise models achieving baseline success rates in the 48-51% range, underscoring the intricacy of the problem. By offering an execution-based benchmark emphasizing the dynamic nature of code libraries, GitChameleon 2.0 enables a clearer understanding of this challenge and helps guide the development of more adaptable and dependable AI code generation methods.

Ultrasound Localization Microscopy (ULM) is a non-invasive technique that allows for the imaging of micro-vessels in vivo, at depth and with a resolution on the order of ten microns. ULM is based on the sub-resolution localization of individual microbubbles injected in the bloodstream. Mapping the whole angioarchitecture requires the accumulation of microbubbles trajectories from thousands of frames, typically acquired over a few minutes. ULM acquisition times can be reduced by increasing the microbubble concentration, but requires more advanced algorithms to detect them individually. Several deep learning approaches have been proposed for this task, but they remain limited to 2D imaging, in part due to the associated large memory requirements. Herein, we propose the use of sparse tensor neural networks to enable deep learning-based 3D ULM by improving memory scalability with increased dimensionality. We study several approaches to efficiently convert ultrasound data into a sparse format and study the impact of the associated loss of information. When applied in 2D, the sparse formulation reduces the memory requirements by a factor 2 at the cost of a small reduction of performance when compared against dense networks. In 3D, the proposed approach reduces memory requirements by two order of magnitude while largely outperforming conventional ULM in high concentration settings. We show that Sparse Tensor Neural Networks in 3D ULM allow for the same benefits as dense deep learning based method in 2D ULM i.e. the use of higher concentration in silico and reduced acquisition time.

Ultrasound Localization Microscopy (ULM) has proven effective in resolving microvascular structures and local mean velocities at sub-diffraction-limited scales, offering high-resolution imaging capabilities. Dynamic ULM (DULM) enables the creation of angiography or velocity movies throughout cardiac cycles. Currently, these techniques rely on a Localization-and-Tracking (LAT) workflow consisting in detecting microbubbles (MB) in the frames before pairing them to generate tracks. While conventional LAT methods perform well at low concentrations, they suffer from longer acquisition times and degraded localization and tracking accuracy at higher concentrations, leading to biased angiogram reconstruction and velocity estimation. In this study, we propose a novel approach to address these challenges by reversing the current workflow. The proposed method, Tracking-and-Localization (TAL), relies on first tracking the MB and then performing localization. Through comprehensive benchmarking using both in silico and in vivo experiments and employing various metrics to quantify ULM angiography and velocity maps, we demonstrate that the TAL method consistently outperforms the reference LAT workflow. Moreover, when applied to DULM, TAL successfully extracts velocity variations along the cardiac cycle with improved repeatability. The findings of this work highlight the effectiveness of the TAL approach in overcoming the limitations of conventional LAT methods, providing enhanced ULM angiography and velocity imaging.

Ultrasound Localization Microscopy (ULM) can map microvessels at a resolution of a few micrometers ( [Formula: see text]). Transcranial ULM remains challenging in presence of aberrations caused by the skull, which lead to localization errors. Herein, we propose a deep learning approach based on recently introduced complex-valued convolutional neural networks (CV-CNNs) to retrieve the aberration function, which can then be used to form enhanced images using standard delay-and-sum beamforming. CV-CNNs were selected as they can apply time delays through multiplication with in-phase quadrature input data. Predicting the aberration function rather than corrected images also confers enhanced explainability to the network. In addition, 3D spatiotemporal convolutions were used for the network to leverage entire microbubble tracks. For training and validation, we used an anatomically and hemodynamically realistic mouse brain microvascular network model to simulate the flow of microbubbles in presence of aberration. The proposed CV-CNN performance was compared to the coherence-based method by using microbubble tracks. We then confirmed the capability of the proposed network to generalize to transcranial in vivo data in the mouse brain (n=3). Vascular reconstructions using a locally predicted aberration function included additional and sharper vessels. The CV-CNN was more robust than the coherence-based method and could perform aberration correction in a 6-month-old mouse. After correction, we measured a resolution of [Formula: see text] for younger mice, representing an improvement of 25.8%, while the resolution was improved by 13.9% for the 6-month-old mouse. This work leads to different applications for complex-valued convolutions in biomedical imaging and strategies to perform transcranial ULM.

Both theoretical and practical problems in deep learning classification require solutions for assessing uncertainty prediction but current state-of-the-art methods in this area are computationally expensive. In this paper, we propose a new confidence measure dubbed Over-MAP that utilizes a measure of overlap between structural attention mechanisms and segmentation methods, that is of particular interest in accurate fine-grained contexts. We show that this classification confidence increases with the degree of overlap. The associated confidence and identification tools are conceptually simple, efficient, and of high practical interest as they allow for weeding out misleading examples in training data. Our measure is currently deployed in the real-world on widely used platforms to annotate large-scale data efficiently.

Both theoretical and practical problems in deep learning classification benefit from assessing uncertainty prediction. In addition, current state-of-the-art methods in this area are computationally expensive: for example,~\cite{loquercio2020general} is a general method for uncertainty estimation in deep learning that relies on Monte-Carlo sampling. We propose a new, efficient confidence measure later dubbed Over-MAP that utilizes a measure of overlap between structural attention mechanisms and segmentation methods. It does not rely on sampling or retraining. We show that the classification confidence increases with the degree of overlap. The associated confidence and identification tools are conceptually simple, efficient and of high practical interest as they allow for weeding out misleading examples in training data. Our measure is currently deployed in the real-world on widely used platforms to annotate large-scale data efficiently.