Foundation models (FMs) and vision-language models (VLMs) have emerged as a transformative paradigm in artificial intelligence, reshaping the landscape of medical image analysis and clinical decision support. Building upon large-scale pretraining on heterogeneous datasets, these models exhibit unprecedented generalization and zero-shot transfer capabilities, effectively overcoming the perennial bottleneck of annotated data scarcity that has historically constrained deep learning applications in healthcare. This review paper presents the state of the art in FMs and VLMs applied to medical imaging, encompassing a broad spectrum of modalities including radiography, computed tomography (CT), magnetic resonance imaging (MRI), pathology whole-slide images, ophthalmology, dermatology, and echocardiography. A structured taxonomy of architectures, training strategies, and adaptation mechanisms is provided, covering contrastive learning (CLIP-based), generative, and hybrid paradigms. Also, seminal works are critically analyzed, such as MedCLIP, BioViL, BiomedCLIP, MedSAM, SAM-Med2D, LLaVA-Med, Med-PaLM 2, CONCH, and BiomedGPT, examining their methodological contributions, benchmark performance, and inherent limitations. Additionally, research gaps are discussed, including insufficient 3D volumetric understanding, limited multimodal reasoning in rare disease contexts, translational barriers for clinical deployment, and critical ethical concerns such as algorithmic bias and privacy. Further, emerging trends are presented encompassing instruction-tuned multimodal large language models (LLMs), agent-based systems, federated learning, and FMs for time-series physiological data. Finally, research directions are highlighted for clinically validated, equitable, and deployable AI systems in medical imaging.
This study proposes a Volterra-inspired time-weighted integral projection (VTIP) for constructing low-dimensional representations of event-level temporal data with internal phase transitions and duration structure. A predefined input template and temporal weighting function yield an event-specific integral factor that is combined with low-order event statistics and their interaction term in a three-dimensional feature vector. The analytical framework examines marginal temporal sensitivity, local second-order organization, and feature-space structure. In 243 saccadic events obtained from six participants, the univariate linear associations varied across features and movement directions. The event duration showed comparatively stronger explanatory power for the amplitude-related component, whereas several switching-moment, velocity-related, and interaction models had limited explanatory power. Peak velocity-related Hessian terms showed high sign consistency under repeated event subsampling, whereas amplitude-related terms were more dependent on direction and sample composition. Relative to the original event-statistical space, the PC1 explained proportion of VTIP increased by approximately 6.4% and 7.7% for horizontal and vertical saccades, respectively. A strict duration scale-matched representation was comparable to or better than VTIP for some PCA criteria, and recording-scenario hold-out analysis did not yield a consistent ranking across movement directions. These findings identify duration scale as a major source of the observed low-dimensional organization and position VTIP as a structured event-level representation that integrates event statistics, duration, and relative switching position. These results characterize the mathematical structure and within-dataset behavior of VTIP. They do not establish participant-level generalizability, external predictive validity, or consistent superiority over simpler duration-matched or trajectory-based representations.
Accurate regulation of the outlet hydrocarbon-vapor temperature in the Alberta Taciuk Process scrubber is challenged by nonlinear coupling, strong thermal inertia, transport delays, and nonstationary operating conditions. This paper develops a residual-adaptive robust preview linear quadratic regulator (RA-RPLQR) that integrates GRU-based process modeling with optimal control. A gated recurrent unit (GRU) trained on industrial time-series data captures nonlinear temporal relationships and provides one-step-ahead temperature predictions without requiring a detailed first-principles model. Automatic differentiation extracts local time-varying Jacobians with respect to the manipulated inputs, connecting the learned dynamics to the LQR design and enabling the control model to be updated as operating conditions change. Measurable upstream disturbances with available lead time are incorporated through preview feedforward compensation. Model uncertainty and unmeasured disturbances are represented through a robust linear-quadratic dynamic-game formulation, in which the unknown-disturbance channel is adjusted online using a recursively filtered GRU prediction residual. Feedback and preview gains are obtained through a finite-horizon generalized Riccati recursion. Actuator magnitude and rate limits are enforced through command-level feasibility projection, together with back-calculation anti-windup compensation during saturation. Closed-loop simulations constructed from previously unseen industrial operating trajectories from the same scrubber show improved tracking, disturbance attenuation, and actuator smoothness compared with PID, conventional LQR, and Adaptive LQR within the evaluated scenarios. The proposed method achieves an integral absolute error of 56.00 and an integral time-weighted absolute error of 1516.39 while reducing transient deviations and actuator movement.
When numerical predictions and ranked recommendations have task value, selection must expose their trade-offs. We examine selection consequences and incremental uncertainty value in biased matrix factorization. Across 35 configurations and four training-evidence levels on MovieLens 1M, CiaoDVD and Yelp–Tampa, validation root mean squared error (RMSE) and normalized discounted cumulative gain for the top ten recommendations (NDCG@10) select different configurations in eleven of twelve conditions. Fresh-seed confirmation shows positive ranking regret whenever selections differ, but only four intervals lie wholly above the declared absolute tolerance of 0.005. An uncertainty-aware (UA) rule minimizes RMSE subject to a paired-seed upper bound on validation ranking regret. A revision-stage mean-regret ablation removes only the uncertainty penalty. At 0.005 tolerance, choices differ in seven of sixteen conditions, and UA improves held-out ranking point estimates in six at higher global RMSE, with no uniform improvement across both endpoints. Four-margin reselection and validation-user resampling expose sensitivity. In the frozen prospective Epinions evaluation, 35 candidates are admissible and UA equals RMSE selection. All four absolute-margin tests pass, yet ranking losses represent 28.2–62.2% of reference NDCG. Expanded searches retain objective disagreement and finite budget limits. The results characterize conditional trade-offs, not universal superiority, so an absolute tolerance can constrain decisions without ensuring proportional ranking preservation or deployment utility.
Identifying super spreaders is a core task in data stream processing. For this purpose, probabilistic data structures called sketches are widely used because they trade memory for estimation accuracy. Many existing sketch-based methods detect super spreaders inside fixed-length intervals known as epochs, clearing the sketch when each epoch ends. However, flushing the sketch in this manner can introduce a transient rebuilding phase during which the accumulated statistics have not yet stabilized. A super spreader that emerges within this interval may therefore be missed. On the other hand, if the stream is measured continuously without ever resetting the sketch, outdated entries build up over time and can reduce accuracy. To address this trade-off, we introduce CycleSketch, a sketch that performs continuous super spreader detection while purging stale data. Concretely, CycleSketch combines a sliding window with a circular queue to keep track of candidate super spreaders and cyclically overwrites aged entries by resetting the slot associated with each new epoch—without globally clearing the count part. To gauge its practical effectiveness, we build and benchmark CycleSketch on both conventional CPU hardware and the OVS virtual switch. On the evaluated traffic traces and under the tested settings, CycleSketch achieves higher measured precision and F1 score than the compared sketch-based detectors, with improvements of 50% and 125%, respectively.
