Design and document ROI/VOI/mask annotation that survives Radiology (RSNA) review — lesion-selection strategy (2D vs 3D, whole-tumour vs largest-slice vs peritumoral vs habitat vs multi-lesion), reader protocol (number, seniority, blinding, independent vs consensus, third-party adjudication), reproducibility (repeat annotation, ICC, Dice, Hausdorff, feature-stability filtering), and geometric integrity of masks against DICOM/NIfTI/DICOM-SEG/RTSTRUCT (spacing, origin, direction, slice order). Use when the user mentions ROI, VOI, mask, segmentation protocol, contouring, "标注", "勾画", "分割规范", inter-/intra-observer agreement, or needs the Methods paragraph for annotation. Produces an annotation SOP, a QC plan, and submission-ready Methods text. Never invents agreement values or reader details.
Turn manuscript text, claims, figure/table statements, abstracts, slides, or novelty/comparison assertions into verified, imaging-journal-scoped citation candidates and export one reference-manager-ready file (RIS, EndNote ENW, or BibTeX). Use when the user needs references for an imaging paper, wants supporting citations, wants a two-pass claim/citation/numerical verification gate, needs to check whether a cited source actually says the claimed thing, scope citations to radiology/imaging journals (Radiology, Radiology: AI, RadioGraphics, AJR, European Radiology, JACR, etc.), verify a DOI/PMID, or export a bibliography. Verifies identifiers before formatting and never fabricates DOIs, pages, volumes, journal metadata, or source support.
Prepare and audit Data/Code Availability statements, DICOM de-identification plans, repository selection, dataset citations, and FAIR/sharing checks for Radiology (RSNA) and Nature-portfolio imaging+omics submissions. Use when the user needs a data availability statement, must de-identify DICOM imaging, choose a repository (TCIA, Zenodo, GEO, dbGaP, EGA), share code/models, write dataset citations, check FAIR compliance, or needs Extended Data / Supplementary Information / Source Data planning for a Nature-family journal — including controlled-access genomics for radiogenomics. Bilingual-aware (中文作者备注 → submission-ready English). Never overstates availability or invents accessions.
Design and audit imaging deep-learning studies to Radiology (RSNA) / CLAIM 2024 standard, or to Nature-portfolio / FUTURE-AI trustworthy-AI standard — architecture choice (2D/2.5D/3D CNN, Transformer/ViT, segmentation/detection nets, prognostic models), transfer learning vs self-supervised pretraining vs training from scratch, how images/masks/clinical/text/molecular inputs enter the model, data splitting and augmentation, class imbalance, hyperparameter search, baselines, external validation, interpretability/explainability (Grad-CAM, SHAP, attention), uncertainty quantification (MC dropout, ensembles, conformal prediction), and robustness/OOD testing — with patient-level partition hygiene throughout. Use when the user plans or reviews a CNN/Transformer/3D/segmentation/detection/foundation/multimodal imaging model, mentions transfer learning, self-supervised, nnU-Net, ViT, data augmentation, class imbalance, explainability, uncertainty, robustness, or "影像深度学习/深度学习模型". Produces a model+training+validation des
Assess whether an imaging dataset can support a study and turn it into a complete, submittable design — from feasibility triage to clinical question, target population, endpoint/estimand, minimum-viable vs stronger methods, and a validation strategy (internal resampling, temporal, geographic, fully external, multi-center, federated). Use when the user has CT/MRI/PET/US/mammography/multimodal data but is unsure what to do, asks "can this topic be done?" / "能不能做" / "帮我设计课题" / "study design" / "what can I study with this data", or needs a multi-center / external-validation plan ("多中心", "external validation", "generalisability", "center effect", "scanner effect"). Produces a study blueprint with feasibility verdict, design options, validation plan, and the limiting constraint surfaced. Never invents cohort numbers or overstates what the data can support.
Write compliant, non-overstated Ethics, Informed-Consent, and privacy/governance text for imaging and imaging+omics studies — study-type framing (retrospective, prospective, registered, multi-center), IRB/ethics-committee approval and numbers, consent vs documented waiver, re-identification risk (DICOM metadata, rare-disease small cohorts, genomic data), and the human-subjects governance behind data sharing (HIPAA/GDPR/PIPL, data-use agreements, controlled access). Use when the user mentions ethics, IRB, ethics committee, informed consent, consent waiver, "伦理", "知情同意", "隐私", HIPAA, GDPR, PIPL, data-use agreement, or needs the Ethics/Consent statement for a manuscript or grant. Produces submission-ready statements and a re-identification risk read. Never fabricates approval numbers, dates, or consent status.
Produce publication-quality white-background academic figures for Radiology (RSNA), Nature-portfolio/npj, European Radiology, NEJM, Science, or Lancet-family venues with Python (matplotlib): ROC curves, calibration plots, decision-curve analysis, forest/SROC plots, Kaplan-Meier curves with numbers-at-risk, Bland-Altman, heatmaps, radiogenomics plots, graphical/visual abstracts, and annotated imaging panels. Uses The Lancet Digital Health guide as the default Lancet-series proxy. Use when the user wants figures, plot cleanup, figure-set planning, journal-specific figure formatting, or overlap/crowding QA. Outputs editable vector (.svg/.pdf) plus 300+ dpi raster, enforces de-identification and journal typography, and never invents data points.
Find publishable frontier directions and innovation points for imaging-AI / radiomics / radiogenomics research, grounded in the publication patterns of high-impact journals (Radiology, Radiology: AI, Lancet Digital Health, Lancet Oncology, Nature Medicine, Nature Communications, npj Digital Medicine, npj Precision Oncology, eClinicalMedicine, Cell Reports Medicine). Use when the user asks for frontier directions, innovation points, hot vs suitable topics, "近三年前沿方向", "创新点", "what's novel in imaging AI", or wants to know the evidence/publication-pattern basis behind a recommendation ("有什么文献依据", "证据"). Translates trends (foundation models, self-supervised, vision-language, multimodal fusion, longitudinal, weak/federated learning, radiogenomics) into executable research questions matched to the user's actual data. Encodes publication-pattern heuristics, not a fabricated citation list — routes live verification to radiology-search and never invents PMIDs/DOIs.