Institution profile

National Institutes of Health

Academic institutionnorthamerica · us
Official website
Research library119linked papers
Opportunities0open roles
Selected work

Representative Papers

Class Imbalance Correction for Improved Universal Lesion Detection and Tagging in CT

Apr 08, 2025MILLanD@MICCAI

The DeepLesion dataset exhibits severe fine-grained class imbalance across anatomical regions, per-patient lesion counts, and lesion sizes, significantly hindering multi-class lesion detection and anatomical localization in CT imaging. To address this, we propose the first multi-dimensional data rebalancing paradigm tailored for joint lesion detection and anatomical localization—built upon the VFNet framework. We systematically compare hierarchical sampling strategies: anatomical region–based, patient-level, and lesion-size–based sampling, and—crucially—first identify and explicitly model DeepLesion’s three-tiered imbalance structure. Additionally, we introduce a standardized radiology report template, specifically structuring the “lesion” subsection to support rigorous data curation. Experiments demonstrate substantial improvements: on lesions ≥1 cm, sensitivity reaches 80% (bone), 77% (kidney), 70% (soft tissue), and 83% (pelvis)—all markedly surpassing random-sampling baselines. Notably, lesion-size balancing further boosts recall across all categories.

4 citationsRead paper

Correcting class imbalances with self-training for improved universal lesion detection and tagging

Apr 07, 2023Medical Imaging

In universal lesion detection (ULDT) from CT scans, the DeepLesion dataset suffers from incomplete annotations and severe class imbalance, hindering robust multi-class lesion detection. Method: We propose a multi-round self-training framework built upon the VFNet detector, incorporating dynamic confidence-thresholding for pseudo-label selection, undersampling-guided oversampling of underrepresented lesion classes, and iterative refinement of pseudo-labels—entirely without additional manual annotation. Contribution/Results: To our knowledge, this is the first method achieving simultaneous sensitivity improvement across all eight lesion classes under a strict 4 false positives per scan (4FP) constraint. The overall sensitivity reaches 78.5%, representing an absolute gain of 11.7% over the baseline. Critically, detection performance does not degrade for any anatomical region; gains are especially pronounced for minority classes. The approach significantly enhances model generalizability and clinical applicability.

2 citationsRead paper

3D universal lesion detection and tagging in CT with self-training

Apr 07, 2023Medical Imaging

3D universal lesion detection in CT volumes faces challenges including absence of voxel-level 3D annotations in DeepLesion, severe class imbalance, and lack of anatomical localization. Method: We propose the first fully automatic framework supporting 3D lesion localization, fine-grained classification, and joint anatomical region annotation. Our approach builds upon VFNet as the 2D detection backbone, introduces a novel 2D→3D contextual expansion mechanism, and employs a multi-round self-training strategy—achieving performance comparable to full supervision using only 30% of DeepLesion data. Results: The method attains a mean sensitivity of 46.9% across 0.125–8 false positives per scan, matching the fully supervised baseline. It is the first to enable end-to-end 3D lesion detection with concurrent anatomical region labeling, significantly enhancing clinical utility and generalizability.

2 citationsRead paper

Reconstructing Sepsis Trajectories from Clinical Case Reports using LLMs: the Textual Time Series Corpus for Sepsis

Apr 12, 2025

This study addresses the lack of structured temporal sequences in sepsis-related clinical text by proposing the first end-to-end large language model (LLM) pipeline to automatically extract and temporally anchor Sepsis-3–defined events from unstructured, non-temporal discharge summaries and case reports. The method integrates prompt engineering with O1-preview and Llama-3.3-70B-Instruct, joint entity–temporal span extraction, rule-augmented post-processing, and a cross-source validation framework (I2B2/MIMIC-IV). It presents the first systematic evaluation of LLMs’ temporal grounding capability in clinical narratives, exposing inherent limitations in pure-text sequential reasoning and identifying multimodal augmentation as a critical future direction. On 2,139 PubMed-open case reports, the pipeline achieves an event matching rate of 0.755 and temporal ordering consistency of 0.932. We release the first publicly available Sepsis-3 textual temporal sequence corpus, enabling dynamic clinical modeling and interpretable AI research.

1 citationsRead paper

Leveraging anatomical priors for automated pancreas segmentation on abdominal CT

Apr 04, 2025Medical Imaging 2025: Computer-Aided Diagnosis

Automatic pancreatic segmentation in abdominal CT suffers from low accuracy and high false-negative rates. Method: We propose a 3D full-resolution nnU-Net framework integrated with anatomical prior knowledge—specifically, multi-organ anatomical labels from TotalSegmentator, leveraging spatial constraints from neighboring organs as strong priors, and jointly trained end-to-end on the PANORAMA dataset. Contribution/Results: Our method significantly improves segmentation robustness: Dice score increases by 6.0% (p < 0.001), Hausdorff distance decreases by 36.5 mm (p < 0.001), and achieves 100% pancreatic detection with zero false negatives. This work empirically validates the critical value of anatomical priors for fine-grained single-organ segmentation, establishing a highly reliable foundation for radiomic biomarker extraction and pancreatic lesion identification.

1 citationsRead paper
Recent publications

Latest Papers