Zeyu Gao

Artefacts compromise clinical decision-making in the use of medical time series. Pulsatile waveforms offer opportunities for accurate artefact detection, yet most approaches rely on supervised manners and overlook patient-level distribution shifts. To address these issues, we introduce GenClean, a generalised label-free framework for real-time artefact cleaning, implemented within the ICM+ clinical research monitoring software. Leveraging an in-house dataset of 180,000 ten-second arterial blood pressure (ABP) samples for training, we first investigate patient-level generalisation, demonstrating robust performance under both intra- and inter-patient distribution shifts. As an initial exploration beyond the development cohort, we further validate its effectiveness for ABP through site-level generalisation on the MIMIC-III database. We also provided an extension of our method to photoplethysmography (PPG), highlighting its potential applicability to diverse medical pulsatile signals. The real-time integration and these generalisation studies collectively demonstrate the practical utility of our framework in continuous physiological monitoring and represent a promising step towards improving the reliability of high-resolution medical time series analysis.

With the widespread adoption of pathology foundation models in research and clinical decision support, their security has become a critical concern, while their vulnerability to adversarial attacks remains largely unexplored. In this work, we present the first systematic investigation into the security of pathology foundation models for whole slide image (WSI) analysis against adversarial attacks. We introduce the principle of local perturbation with global impact and develop an attack framework under a sparse, task-agnostic, white-box threat model. Considering the gigapixel and patch-based characteristics of pathology images, we propagate pixel-level perturbations through selected patches to induce global effects at the WSI level. Based on this principle, we adapt four classical white-box attack methods to the task-agnostic WSI setting and redefine the perturbation budget. We establish a cross-model, multi-dataset adversarial benchmark covering five pathology foundation models, five datasets, and eight downstream tasks under a unified evaluation protocol, assessing performance degradation, attack success, perturbation sensitivity, and defence effectiveness. Our results show that perturbing only 0.1% of patches per slide with imperceptible noise can substantially degrade downstream performance, with performance drops exceeding 60% in the most vulnerable cases. We further evaluate lightweight input-transformation defences and analyze factors affecting attack effectiveness, including perturbation budget, multi-patch synergy, and objective design. Finally, we provide mechanistic evidence that patch-level vulnerability is associated with semantic content, while adversarial perturbations shift attention toward attacked regions and may amplify local noise into slide-level representation changes.

Large Vision Language Models (LVLMs) are increasingly used in computational pathology for image classification, description generation, question answering and interactive diagnostics. However, most pathology LVLMs analyse small regions of interest rather than pyramidal, gigapixel-scale whole-slide images (WSIs), limiting their use for tasks requiring whole-slide assessment across sub-regions and magnification levels. Here, we present ALPaCA (Adapting Llama for Pathology Context Analysis), a slide-level LVLM framework for WSI question answering across diverse cancer types and tissue sites. ALPaCA is trained using 35,913 WSIs with curated descriptions and 341,051 question-answer pairs from TCGA and GTEx. It combines a LongFormer vision-text adaptor with a Gaussian mixture model-based prototyping adaptor and Llama3.1. ALPaCA exceeds 90% accuracy on internal close-ended benchmarks and maintains 77–82% accuracy on independent external cohorts. Expert pathologist evaluation of open-ended responses supports its slide-level reasoning capability. Additionally, ALPaCA can be fine-tuned on organ- or disease-specific datasets, supporting specialised pathology question answering.

Motivation: Whole-slide image (WSI)-based biomarker prediction in computational pathology has the potential to support scalable and resource-efficient analysis of large pathology cohorts, helping prioritize cases for downstream molecular testing and patient stratification. However, reliable biomarker labels are often limited and costly to obtain, making label-efficient WSI prediction essential. Recent slide-level foundation models have opened new opportunities for few-shot biomarker prediction by providing strong pretrained slide representations. Nevertheless, few-shot learning still suffers from sparse labeled support data, and data augmentation remains important for improving robustness and generalization. In frozen multi-stage WSI pipelines, however, conventional augmentation is difficult to apply: pixel-level augmentation requires costly feature re-extraction, while naive perturbation of frozen representations may compromise semantic consistency. Results: To address this challenge, we propose Learnable Frozen Feature Augmentation (LFFA), a training-time feature-space augmentation framework for few-shot WSI biomarker prediction. Instead of directly perturbing frozen slide features, LFFA learns controllable augmented slide views from contextualized interaction tokens under geometry-aware and downstream-supervised constraints, improving representation diversity while preserving semantic consistency. The augmented features are further optimized with an augmented-class α-mix loss to balance diversity and class semantics. We evaluate LFFA on three few-shot WSI biomarker prediction tasks, covering molecular marker prediction and gene mutation prediction, using three representative and widely used slide-level foundation models. Results show that LFFA consistently improves the corresponding baselines and achieves stronger overall robustness than existing augmentation methods across tasks, backbones, and shot settings. Availability and implementation: Source code and processed experimental splits will be made available at https://github.com/zdipath/LFFA.