Chronological feed of everything captured from Andrew Curran.
DyABD is the first benchmark dataset for abdominal muscle segmentation in dynamic MRI, featuring patients with abdominal hernias performing exercises — introducing extreme anatomical variability rare in medical imaging benchmarks. The dataset includes pre- and post-operative MRIs and is designed to advance clinical understanding of hernia recurrence. A comprehensive evaluation across supervised, few-shot, and zero-shot segmentation paradigms reveals that current state-of-the-art models largely plateau around a Dice Coefficient of 0.82 on unseen DyABD data, indicating meaningful headroom for improvement in medical image segmentation generalization.
New ³⁶Cl measurements from multiple polar ice core sites (EGRIP and NGRIP in Greenland, and Law Dome in Antarctica), combined with semi-annual ¹⁰Be data, show no significant radionuclide spike around 1859 CE. This absence of cosmogenic signal allows researchers to rule out an extreme solar energetic particle (SEP) event associated with the Carrington geomagnetic storm, fundamentally challenging its use as a worst-case SEP benchmark. The study narrows the Carrington event to two scenarios: either a relatively soft SEP event (fluence above 30 MeV up to 3× the largest Space Age event), or no Earth-directed SEP event at all.
ABRA is a new radiology-agent benchmark that evaluates LLM agents operating a real OHIF viewer and Orthanc DICOM server via 21 function-calling tools, across 655 programmatically generated tasks spanning navigation, annotation, and structured reporting. Critically, ABRA decouples tool orchestration from visual perception by pairing each task with an "oracle" variant where a simulated detector supplies the finding. Results across 10 models (5 closed-, 5 open-weight) show high execution competence (≥89%) but near-zero task outcomes (0–25%), with oracle variants achieving 69–100% outcome — conclusively localizing the failure mode to medical image perception rather than agentic planning or tool use.