Institution profile

Harvard Medical School

Academic institutionnorthamerica · us
Official website
Research library609linked papers
Opportunities0open roles
Selected work

Representative Papers

Medical Hallucinations in Foundation Models and Their Impact on Healthcare

Feb 26, 2025arXiv.org

Medical foundation models may generate “hallucinations”—factual, logical, or evidence-inconsistent errors—that jeopardize clinical decision-making and patient safety. To address this, we first propose a multidimensional taxonomy of medical hallucinations and establish a real-world, clinician-annotated benchmark dataset derived from authentic clinical cases; we further validate its clinical impact via an international physician survey. Methodologically, we integrate expert annotation, empirical behavioral surveys, and large language model (LLM) evaluation to systematically assess the efficacy of chain-of-thought (CoT) reasoning and retrieval-augmented generation (RAG) in mitigating hallucinations. Results show both techniques significantly reduce hallucination rates, yet residual hallucinations remain clinically hazardous. Building on these findings, we introduce a patient-safety-centered AI governance and ethics framework, offering theoretical foundations and actionable pathways for responsible deployment of medical AI. (149 words)

41 citationsRead paper

Benchmarking the CoW with the TopCoW Challenge: Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA

Dec 29, 2023arXiv.org

The Circle of Willis (CoW) suffers from a scarcity of high-quality voxel-level annotations in CTA/MRA imaging, reliance on labor-intensive expert manual segmentation, and poor guarantee of topological consistency. Method: We introduce the first publicly available voxel-level multi-class CoW dataset—comprising 13 vascular structures with paired MRA/CTA volumes—and propose a topology-aware segmentation framework: (i) a novel VR-assisted annotation paradigm ensuring anatomical plausibility; (ii) a multimodal registration and topology-constrained segmentation network; and (iii) topology-sensitive metrics including branch F1 and topo-Dice. Contribution/Results: This benchmark has attracted >140 teams across four continents. State-of-the-art models achieve ≈90% Dice on most arterial branches, while exposing persistent topological matching bottlenecks—particularly for communicating arteries and anatomical variants.

24 citations2 influentialRead paper

Diffusion MRI with machine learning

Jan 01, 2024Imaging Neuroscience

dMRI analysis faces core challenges including severe noise, high inter-scanner and inter-subject variability, and complex microstructural modeling. This study systematically reviews and empirically evaluates machine learning across the full dMRI pipeline—signal denoising, harmonization, microstructural mapping, fiber tractography, and white-matter pathway quantification—and establishes, for the first time, its applicability boundaries. We innovatively integrate CNNs, GANs, VAEs, transfer learning, multi-site harmonization, and explainable AI (XAI), while proposing a benchmark dataset construction and validation framework tailored for clinical deployment. Our analysis identifies shared bottlenecks in model robustness, reproducibility, and interpretability, and identifies generalizability and standardized evaluation as critical leverage points. The work provides a methodological guide and research roadmap for developing trustworthy, reproducible, and interpretable next-generation dMRI analysis tools.

14 citationsRead paper

End-to-end deep learning for interior tomography with low-dose x-ray CT

Apr 07, 2022Physics in Medicine and Biology

To address the strong coupling between cupping artifacts and quantum noise caused by truncated projections in low-dose X-ray CT, this paper proposes a dual-domain end-to-end deep learning framework: denoising in the image domain and truncation-aware projection data extrapolation in the sinogram domain, synergistically enabling high-fidelity interior reconstruction. We introduce the novel “dual-domain decoupled modeling” paradigm, overcoming the fundamental limitation of single-domain CNNs in disentangling coupled artifacts. To our knowledge, this is the first work to demonstrate that sinogram-domain CNNs outperform state-of-the-art image-domain methods under combined truncation and low-dose conditions. The network architecture is theoretically grounded in deep convolutional principles and jointly optimizes two parallel branches. Experiments show significant improvements over image-domain SOTA methods in PSNR and SSIM; sinogram-domain reconstruction accuracy increases by over 15%; cupping artifacts and noise are effectively suppressed.

10 citationsRead paper
Recent publications

Latest Papers