Insights

Notes on why the constraint in cell modelling moved from compute to data format, how an annotation tool accumulated a corpus mostly contributed by others, and what rare-disease groups bring beyond motivation.

Notes on why biology's bottleneck is data that does not exist yet, what it means that protein design arrived without antibody-specific model development, and the argument that interpretability could yield biological knowledge.

Notes on whether the image causes the label or the label causes the image, how that framing bears on data scarcity and dataset shift, and an experiment in which site information survives a full neuroimaging pipeline.

Notes on SHAP’s axioms, image ablation, and the Prescience clinician study, with implications for evaluating clinical faithfulness audits of medical image classifiers.

Notes on Prescience, ENABL Age, Alzheimer’s pathway models, COVID shortcuts, and CoAI, with implications for clinical faithfulness auditing.

Notes from a TWIML conversation on SHAP, biological interpretation, counterfactual clinical audits, and the connection to my research on clinical faithfulness in gallbladder ultrasound AI.

Notes from Prof. Taeho Jo on using AI across tau PET, whole-genome sequencing, metabolomics, rare variants, and uncertainty estimation to search for earlier and more trustworthy signals of Alzheimer's disease.

Notes on UDCD: mean-teacher self-distillation, contrastive relation matrices, entropy-based confidence weighting, and why a medical-AI student should not imitate every signal from its teacher equally.

Notes from a lecture by Prof. Joon-Beom Seo, head of the AI-Basic Healthcare Task Force: ten years from AlphaGo, four ways AI can transform care, and the case for a sovereign medical foundation model.

Notes on Korea's new national medical AI strategy: outcome-based reimbursement decided at the hospital level, a public 'AI highway' for inference infrastructure, and sovereign medical AI.

Notes from Ricardo Gonzales's RISE-MICCAI talk on ROADMAP, RSNA ATLAS, machine-readable performance metrics, LLM-assisted documentation, and why transparency can expose, but not eliminate, gaps in medical AI evidence.

Notes from former MFDS reviewer Dr. Chungkeun Lee on how Korea moved early in medical-AI regulation, why model changes matter as much as initial accuracy, and how generative AI is pushing regulation from one-time approval toward continuous oversight.

Notes from NAVER Cloud's Haanju Yoo on medical LLMs, hallucination, self-consistency, regulation, and why the path to useful healthcare AI begins with defining what can actually be measured.

Notes from Woong Bae's interview on why radiology AI should move beyond lesion detection, how KARA-CXR was designed to draft chest X-ray reports, and why rare but absurd errors matter more than a strong average score.

How MEDBind aligns chest X-rays, ECGs, and clinical text to support multimodal prediction, few-shot learning, and cross-modal knowledge transfer.

Notes from Prof. Chris McIntosh's lecture on automated radiotherapy planning, treatment-specific patient matching, prospective deployment, and the gap between an AI plan being clinically acceptable and clinicians actually using it.

Notes from an interview with Prof. Chris McIntosh on clinical collaboration, radiotherapy deployment, transfer learning, explainability, fairness, wearables, and the long path from a paper to patient care.

Notes from Lunit co-founder Anthony Seungwook Paek on Volpara, foundation models, autonomous AI, and the shift from accurate algorithms to an integrated cancer-care platform.

Notes from an interview with VUNO founder and CEO Ye Ha Lee on DeepCARS, clinical evidence, reimbursement, and what it takes to move medical AI from a model into routine care.

Notes from a MICCAI webinar on regulatory-driven AI development, clinically meaningful endpoints, data quality, post-market drift, SHIFT, and why diagnosis should come before model retraining.

Notes from a MICCAI webinar on performance uncertainty: confidence intervals, false claims of outperformance, data splitting, standard deviation versus standard error, and bootstrap-based reporting.

High average accuracy does not mean medical AI works equally well for every patient. A look at hidden bias, data representation, subgroup performance, and why trustworthy AI must ask who benefits, and who bears the errors.