arXiv subjects

Dingyu Wang

Publications and source records attributed to Dingyu Wang.

14 recordsLinked to original sources

arXiv2026

OrthoDiffusion: A Generalizable Multi-Task Diffusion Foundation Model for Musculoskeletal MRI Interpretation

Musculoskeletal disorders represent a significant global health burden and are a leading cause of disability worldwide. While MRI is essential for accurate diagnosis, its interpretation remains exceptionally challenging. Radiologists must identify multiple potential abnormalities within complex anatomical structures across different imaging planes, a process that requires significant expertise and is prone to variability. We developed OrthoDiffusion, a unified diffusion-based foundation model designed for multi-task musculoskeletal MRI interpretation. The framework utilizes three orientation-specific 3D diffusion models, pre-trained in a self-supervised manner on 15,948 unlabeled knee MRI scans, to learn robust anatomical features from sagittal, coronal, and axial views. These view-specific representations are integrated to support diverse clinical tasks, including anatomical segmentation and multi-label diagnosis. Our evaluation demonstrates that OrthoDiffusion achieves excellent performance in the segmentation of 11 knee structures and the detection of 8 knee abnormalities. The model exhibited remarkable robustness across different clinical centers and MRI field strengths, consistently outperforming traditional supervised models. Notably, in settings where labeled data was scarce, OrthoDiffusion maintained high diagnostic precision using only 10\% of training labels. Furthermore, the anatomical representations learned from knee imaging proved highly transferable to other joints, achieving strong diagnostic performance across 11 diseases of the ankle and shoulder. These findings suggest that diffusion-based foundation models can serve as a unified platform for multi-disease diagnosis and anatomical segmentation, potentially improving the efficiency and accuracy of musculoskeletal MRI interpretation in real-world clinical workflows.

cs.CV↗

arXiv2026

A multimodal vision foundation model for generalizable knee pathology

Musculoskeletal disorders represent a leading cause of global disability, creating an urgent demand for precise interpretation of medical imaging. Current artificial intelligence (AI) approaches in orthopedics predominantly rely on task-specific, supervised learning paradigms. These methods are inherently fragmented, require extensive annotated datasets, and often lack generalizability across different modalities and clinical scenarios. The development of foundation models in this field has been constrained by the scarcity of large-scale, curated, and open-source musculoskeletal datasets. To address these challenges, we introduce OrthoFoundation, a multimodal vision foundation model optimized for musculoskeletal pathology. We constructed a pre-training dataset of 1.2 million unlabeled knee X-ray and MRI images from internal and public databases. Utilizing a Dinov3 backbone, the model was trained via self-supervised contrastive learning to capture robust radiological representations. OrthoFoundation achieves state-of-the-art (SOTA) performance across 14 downstream tasks. It attained superior accuracy in X-ray osteoarthritis diagnosis and ranked first in MRI structural injury detection. The model demonstrated remarkable label efficiency, matching supervised baselines using only 50% of labeled data. Furthermore, despite being pre-trained on knee images, OrthoFoundation exhibited exceptional cross-anatomy generalization to the hip, shoulder, and ankle. OrthoFoundation represents a significant advancement toward general-purpose AI for musculoskeletal imaging. By learning fundamental, joint-agnostic radiological semantics from large-scale multimodal data, it overcomes the limitations of conventional models, which provides a robust framework for reducing annotation burdens and enhancing diagnostic accuracy in clinical practice.

cs.CV↗

arXiv2025

The Illusion of Clinical Reasoning: A Benchmark Reveals the Pervasive Gap in Vision-Language Models for Clinical Competency

Background: The rapid integration of foundation models into clinical practice and public health necessitates a rigorous evaluation of their true clinical reasoning capabilities beyond narrow examination success. Current benchmarks, typically based on medical licensing exams or curated vignettes, fail to capture the integrated, multimodal reasoning essential for real-world patient care. Methods: We developed the Bones and Joints (B&J) Benchmark, a comprehensive evaluation framework comprising 1,245 questions derived from real-world patient cases in orthopedics and sports medicine. This benchmark assesses models across 7 tasks that mirror the clinical reasoning pathway, including knowledge recall, text and image interpretation, diagnosis generation, treatment planning, and rationale provision. We evaluated eleven vision-language models (VLMs) and six large language models (LLMs), comparing their performance against expert-derived ground truth. Results: Our results demonstrate a pronounced performance gap between task types. While state-of-the-art models achieved high accuracy, exceeding 90%, on structured multiple-choice questions, their performance markedly declined on open-ended tasks requiring multimodal integration, with accuracy scarcely reaching 60%. VLMs demonstrated substantial limitations in interpreting medical images and frequently exhibited severe text-driven hallucinations, often ignoring contradictory visual evidence. Notably, models specifically fine-tuned for medical applications showed no consistent advantage over general-purpose counterparts. Conclusions: Current artificial intelligence models are not yet clinically competent for complex, multimodal reasoning. Their safe deployment should currently be limited to supportive, text-based roles. Future advancement in core clinical tasks awaits fundamental breakthroughs in multimodal integration and visual understanding.

cs.CV↗

arXiv2025

Which2comm: An Efficient Collaborative Perception Framework for 3D Object Detection

Collaborative perception allows real-time inter-agent information exchange and thus offers invaluable opportunities to enhance the perception capabilities of individual agents. However, limited communication bandwidth in practical scenarios restricts the inter-agent data transmission volume, consequently resulting in performance declines in collaborative perception systems. This implies a trade-off between perception performance and communication cost. To address this issue, we propose Which2comm, a novel multi-agent 3D object detection framework leveraging object-level sparse features. By integrating semantic information of objects into 3D object detection boxes, we introduce semantic detection boxes (SemDBs). Innovatively transmitting these information-rich object-level sparse features among agents not only significantly reduces the demanding communication volume, but also improves 3D object detection performance. Specifically, a fully sparse network is constructed to extract SemDBs from individual agents; a temporal fusion approach with a relative temporal encoding mechanism is utilized to obtain the comprehensive spatiotemporal features. Extensive experiments on the V2XSet and OPV2V datasets demonstrate that Which2comm consistently outperforms other state-of-the-art methods on both perception performance and communication cost, exhibiting better robustness to real-world latency. These results present that for multi-agent collaborative 3D object detection, transmitting only object-level sparse features is sufficient to achieve high-precision and robust performance.

cs.CV↗

arXiv2024

Multi-dimensional Approximate Counting

The celebrated Morris counter uses $\log_2\log_2 n + O(\log_2 \sigma^{-1})$ bits to count up to $n$ with a relative error $\sigma$, where if $\hat{\lambda}$ is the estimate of the current count $\lambda$, then $\mathbb{E}|\hat{\lambda}-\lambda|^2 <\sigma^2\lambda^2$. A natural generalization is \emph{multi-dimensional} approximate counting. Let $d\geq 1$ be the dimension. The count vector $x\in \mathbb{N}^d$ is incremented entry-wisely over a stream of coordinates $(w_1,\ldots,w_n)\in [d]^n$, where upon receiving $w_k\in[d]$, $x_{w_k}\gets x_{w_k}+1$. A \emph{$d$-dimensional approximate counter} is required to count $d$ coordinates simultaneously and return an estimate $\hat{x}$ of the count vector $x$. Aden-Ali, Han, Nelson, and Yu \cite{aden2022amortized} showed that the trivial solution of using $d$ Morris counters that track $d$ coordinates separately is already optimal in space, \emph{if each entry only allows error relative to itself}, i.e., $\mathbb{E}|\hat{x}_j-x_j|^2<\sigma^2|x_j|^2$ for each $j\in [d]$. However, for another natural error metric -- the \emph{Euclidean mean squared error} $\mathbb{E} |\hat{x}-x|^2$ -- we show that using $d$ separate Morris counters is sub-optimal. In this work, we present a simple and optimal $d$-dimensional counter with Euclidean relative error $\sigma$, i.e., $\mathbb{E} |\hat{x}-x|^2 <\sigma^2|x|^2$ where $|x|=\sqrt{\sum_{j=1}^d x_j^2}$, with a matching lower bound. The upper and lower bounds are proved with ideas that are strikingly simple. The upper bound is constructed with a certain variable-length integer encoding and the lower bound is derived from a straightforward volumetric estimation of sphere covering.

cs.DS↗

arXiv2024

A Unified Construction of Streaming Sketches via the L\'evy-Khintchine Representation Theorem

In the $d$-dimensional turnstile streaming model, a frequency vector $\mathbf{x}=(\mathbf{x}(1),\ldots,\mathbf{x}(n))\in (\mathbb{R}^d)^n$ is updated entry-wisely over a stream. We consider the problem of $f$-moment estimation, where one wants to estimate $$f(\mathbf{x})=\sum_{v\in[n]}f(\mathbf{x}(v))$$ with a small-space sketch. In this work we present a simple and generic scheme to construct sketches with the novel idea of hashing indices to L\'evy processes, from which one can estimate the $f$-moment $f(\mathbf{x})$ where $f$ is the characteristic exponent of the L\'evy process. The fundamental L\'evy-Khintchine representation theorem completely characterizes the space of all possible characteristic exponents, which in turn characterizes the set of $f$-moments that can be estimated by this generic scheme. The new scheme has strong explanatory power. It unifies the construction of many existing sketches and it implies the tractability of many nearly periodic functions that were previously unclassified. Furthermore, the scheme can be conveniently generalized to multidimensional cases ($d\geq 2$) by considering multidimensional L\'evy processes and can be further generalized to estimate heterogeneous moments by projecting different indices with different L\'evy processes. We conjecture that the set of tractable functions can be characterized using the L\'evy-Khintchine representation theorem via what we called the Fourier-Hahn-L\'evy method.

cs.DS↗

arXiv2024

CVVLSNet: Vehicle Location and Speed Estimation Using Partial Connected Vehicle Trajectory Data

Real-time estimation of vehicle locations and speeds is crucial for developing many beneficial transportation applications in traffic management and control, e.g., adaptive signal control. Recent advances in communication technologies facilitate the emergence of connected vehicles (CVs), which can share traffic information with nearby CVs or infrastructures. At the early stage of connectivity, only a portion of vehicles are CVs. The locations and speeds for those non-CVs (NCs) are not accessible and must be estimated to obtain the full traffic information. To address the above problem, this paper proposes a novel CV-based Vehicle Location and Speed estimation network, CVVLSNet, to simultaneously estimate the vehicle locations and speeds exclusively using partial CV trajectory data. A road cell occupancy (RCO) method is first proposed to represent the variable vehicle state information. Spatiotemporal interactions can be integrated by simply fusing the RCO representations. Then, CVVLSNet, taking the Coding-RAte TransformEr (CRATE) network as a backbone, is introduced to estimate the vehicle locations and speeds. Moreover, physical vehicle size constraints are also considered in loss functions. Extensive experiments indicate that the proposed method significantly outperformed the existing method under various CV penetration rates, signal timings, and volume-to-capacity ratios.

cs.CV↗

arXiv2024

Optimal Protocols for 2-Party Contention Resolution

\emph{Contention Resolution} is a fundamental symmetry-breaking problem in which $n$ devices must acquire temporary and exclusive access to some \emph{shared resource}, without the assistance of a mediating authority. For example, the $n$ devices may be sensors that each need to transmit a single packet of data over a broadcast channel. In each time step, devices can (probabilistically) choose to acquire the resource or remain idle; if exactly one device attempts to acquire it, it succeeds, and if two or more devices make an attempt, none succeeds. The complexity of the problem depends heavily on what types of \emph{collision detection} are available. In this paper we consider \emph{acknowledgement-based protocols}, in which devices \underline{only} learn whether their own attempt succeeded or failed; they receive no other feedback from the environment whatsoever, i.e., whether other devices attempted to acquire the resource, succeeded, or failed. Nearly all work on the Contention Resolution problem evaluated the performance of algorithms \emph{asymptotically}, as $n\rightarrow \infty$. In this work we focus on the simplest case of $n=2$ devices, but look for \underline{\emph{precisely}} optimal algorithms. We design provably optimal algorithms under three natural cost metrics: minimizing the expected average of the waiting times ({\sc avg}), the expected waiting time until the first device acquires the resource ({\sc min}), and the expected time until the last device acquires the resource ({\sc max}).

cs.DS↗

arXiv2024

Universal Perfect Samplers for Incremental Streams

If $G : \mathbb{R}_+ \to \mathbb{R}_+$, the $G$-moment of a vector $\mathbf{x}\in\mathbb{R}_+^n$ is $G(\mathbf{x}) = \sum_{v\in[n]} G(\mathbf{x}(v))$ and the $G$-sampling problem is to select an index $v_*\in [n]$ according to its contribution to the $G$-moment, i.e., such that $\Pr(v_*=v) = G(\mathbf{x}(v))/G(\mathbf{x})$. Approximate $G$-samplers may introduce multiplicative and/or additive errors to this probability, and some have a non-trivial probability of failure. In this paper we focus on the exact $G$-sampling problem, where $G$ is selected from the class $\mathcal{G}$ of Laplace exponents of non-negative, one-dimensional L\'evy processes, which includes several well studied classes such as $p$th moments $G(z)=z^p$, $p\in[0,1]$, logarithms $G(z)=\log(1+z)$, Cohen and Geri's soft concave sublinear functions, which are used to approximate concave sublinear functions, including cap statistics. We develop $G$-samplers for a vector $\mathbf{x} \in \mathbb{R}_+^n$ that is presented as an incremental stream of positive updates. In particular: * For any $G\in\mathcal{G}$, we give a very simple $G$-sampler that uses 2 words of memory and stores at all times a $v_*\in [n]$, such that $\Pr(v_*=v)$ is exactly $G(\mathbf{x}(v))/G(\mathbf{x})$. * We give a ``universal'' $\mathcal{G}$-sampler that uses $O(\log n)$ words of memory w.h.p., and given any $G\in \mathcal{G}$ at query time, produces an exact $G$-sample. With an overhead of a factor of $k$, both samplers can be used to $G$-sample a sequence of $k$ indices with or without replacement. Our sampling framework is simple and versatile, and can easily be generalized to sampling from more complex objects like graphs and hypergraphs.

cs.DS↗

arXiv2024

Harmonic Decomposition in Data Sketches

In the turnstile streaming model, a dynamic vector $\mathbf{x}=(\mathbf{x}_1,\ldots,\mathbf{x}_n)\in \mathbb{Z}^n$ is updated by a stream of entry-wise increments/decrements. Let $f\colon\mathbb{Z}\to \mathbb{R}_+$ be a symmetric function with $f(0)=0$. The \emph{$f$-moment} of $\mathbf{x}$ is defined to be $f(\mathbf{x}) := \sum_{v\in[n]}f(\mathbf{x}_v)$. We revisit the problem of constructing a \emph{universal sketch} that can estimate many different $f$-moments. Previous constructions of universal sketches rely on the technique of sampling with respect to the $L_0$-mass (uniform samples) or $L_2$-mass ($L_2$-heavy-hitters), whose universality comes from being able to evaluate the function $f$ over the samples. In this work we take a new approach to constructing a universal sketch that does not use \emph{any} explicit samples but relies on the \emph{harmonic structure} of the target function $f$. The new sketch ($\textsf{SymmetricPoissonTower}$) \emph{embraces} hash collisions instead of avoiding them, which saves multiple $\log n$ factors in space, e.g., when estimating all $L_p$-moments ($f(z) = |z|^p,p\in[0,2]$). For many nearly periodic functions, the new sketch is \emph{exponentially} more efficient than sampling-based methods. We conjecture that the $\textsf{SymmetricPoissonTower}$ sketch is \emph{the} universal sketch that can estimate every tractable function $f$.

cs.DS↗

arXiv2023

Probabilistic Counting in Generalized Turnstile Models

Traditionally in the turnstile model of data streams, there is a state vector $x=(x_1,x_2,\ldots,x_n)$ which is updated through a stream of pairs $(i,k)$ where $i\in [n]$ and $k\in \Z$. Upon receiving $(i,k)$, $x_i\gets x_i + k$. A distinct count algorithm in the turnstile model takes one pass of the stream and then estimates $\norm{x}_0 = |\{i\in[n]\mid x_i\neq 0\}|$ (aka $L_0$, the Hamming norm). In this paper, we define a finite-field version of the turnstile model. Let $F$ be any finite field. Then in the $F$-turnstile model, for each $i\in [n]$, $x_i\in F$; for each update $(i,k)$, $k\in F$. The update $x_i\gets x_i+k$ is then computed in the field $F$. A distinct count algorithm in the $F$-turnstile model takes one pass of the stream and estimates $\norm{x}_{0;F} = |\{i\in[n]\mid x_i\neq 0_F\}|$. We present a simple distinct count algorithm, called $F$-\pcsa{}, in the $F$-turnstile model for any finite field $F$. The new $F$-\pcsa{} algorithm takes $m\log(n)\log (|F|)$ bits of memory and estimates $\norm{x}_{0;F}$ with $O(\frac{1}{\sqrt{m}})$ relative error where the hidden constant depends on the order of the field. $F$-\pcsa{} is straightforward to implement and has several applications in the real world with different choices of $F$. Most notably, it makes distinct count with deletions as simple as distinct count without deletions.

cs.DS↗

arXiv2022

Simpler and Better Cardinality Estimators for HyperLogLog and PCSA

\emph{Cardinality Estimation} (aka \emph{Distinct Elements}) is a classic problem in sketching with many industrial applications. Although sketching \emph{algorithms} are fairly simple, analyzing the cardinality \emph{estimators} is notoriously difficult, and even today the state-of-the-art sketches such as HyperLogLog and (compressed) \PCSA{} are not covered in graduate level Big Data courses. In this paper we define a class of \emph{generalized remaining area} (\tGRA) estimators, and observe that HyperLogLog, LogLog, and some estimators for PCSA are merely instantiations of \tGRA{} for various integral values of $τ$. We then analyze the limiting relative variance of \tGRA{} estimators. It turns out that the standard estimators for HyperLogLog and PCSA can be improved by choosing a \emph{fractional} value of $τ$. The resulting estimators come \emph{very} close to the Cramér-Rao lower bounds for HyperLogLog{} and PCSA derived from their Fisher information. Although the Cramér-Rao lower bound \emph{can} be achieved with the Maximum Likelihood Estimator (MLE), the MLE is cumbersome to compute and dynamically update. In contrast, \tGRA{} estimators are trivial to update in constant time. Our presentation assumes only basic calculus and probability, not any complex analysis~\cite{FlajoletM85,DurandF03,FlajoletFGM07}.

cs.DS↗

arXiv2021

Non-Mergeable Sketching for Cardinality Estimation

Cardinality estimation is perhaps the simplest non-trivial statistical problem that can be solved via sketching. Industrially-deployed sketches like HyperLogLog, MinHash, and PCSA are mergeable, which means that large data sets can be sketched in a distributed environment, and then merged into a single sketch of the whole data set. In the last decade a variety of sketches have been developed that are non-mergeable, but attractive for other reasons. They are simpler, their cardinality estimates are strictly unbiased, and they have substantially lower variance. We evaluate sketching schemes on a reasonably level playing field, in terms of their memory-variance product (MVP). E.g., a sketch that occupies $5m$ bits and whose relative variance is $2/m$ (standard error $\sqrt{2/m}$) has an MVP of $10$. Our contributions are as follows. Cohen and Ting independently discovered what we call the Martingale transform for converting a mergeable sketch into a non-mergeable sketch. We present a simpler way to analyze the limiting MVP of Martingale-type sketches. We prove that the \Martingale{} transform is optimal in the non-mergeable world, and that \Martingale{} \fishmonger{} in particular is optimal among linearizable sketches, with an MVP of $H_0/2 \approx 1.63$. E.g., this is circumstantial evidence that to achieve 1\% standard error, we cannot do better than a 2 kilobyte sketch. \Martingale{} \fishmonger{} is neither simple nor practical. We develop a new mergeable sketch called \Curtain{} that strikes a nice balance between simplicity and efficiency, and prove that \Martingale{} \Curtain{} has limiting $\MVP\approx 2.31$. It can be updated with $O(1)$ memory accesses and it has lower empirical variance than \Martingale{} \LogLog, a practical non-mergeable version of HyperLogLog.

cs.DS↗

arXiv2020

Information Theoretic Limits of Cardinality Estimation: Fisher Meets Shannon

Estimating the cardinality (number of distinct elements) of a large multiset is a classic problem in streaming and sketching. In this paper we study the intrinsic tradeoff between the space complexity of the sketch and its estimation error. We define a new measure of efficiency for data sketches called the Fisher-Shannon (FiSh) number $\mathcal{H}/\mathcal{I}$. It captures the tension between the limiting Shannon entropy ($\mathcal{H}$) of the sketch and its normalized Fisher information ($\mathcal{I}$) that characterizes the variance of a statistically efficient, asymptotically unbiased estimator. Our aim in introducing the FiSh-number is to build the mathematical machinery necessary to argue for precise optimality, rather than asymptotic optimality, up to large constant factors. Our results are as follows. [1] We prove that all base-$q$ variants of Flajolet and Martin's PCSA sketch have FiSh-number $H_0/I_0 \approx 1.98016$ and that every base-$q$ variant of HyperLogLog has FiSh-number worse than $H_0/I_0$, but that they tend to $H_0/I_0$ in the limit as $q\rightarrow \infty$. Here $H_0,I_0$ are precisely defined constants. [2] We describe a sketch called Fishmonger that is based on a smoothed, entropy-compressed variant of PCSA with a different estimator function. Fishmonger processes a multiset of $[U]$ such that at all times, w.h.p., its space is $(1+o(1))(H_0/I_0)m \approx 1.98m$ bits and its standard error is $1/\sqrt{m}$. For example, to achieve a 1% standard error, one needs a little more than 19,800 bits, or $\approx 2.42$ kilobytes. [3] Finally, we give circumstantial evidence that $H_0/I_0$ is the optimum FiSh-number of mergeable sketches for Cardinality Estimation. We define a natural subset of mergeable sketches called linearizable sketches and prove that no member of this class can beat $H_0/I_0$. The popular mergeable sketches are, in fact, also linearizable.

cs.DS↗