MedGemma 1.5 4B integrates high-dimensional CT and MRI volumes via long-context 3D slicing, histopathology whole-slide sampling, bounding-box localization, and multi-timepoint chest X-ray processing into a single Gemma-derived architecture (arXiv:2604.05081). The technical report records an 11 percentage point increase in 3D MRI condition classification, 3 points in 3D CT, 47-point macro F1 gain on whole-slide pathology, 35% IoU improvement on anatomical localization, and 4% macro accuracy lift on longitudinal chest X-rays relative to MedGemma 1 4B. Text-only gains include 5 points on MedQA and 22 points on EHRQA plus 18% average macro F1 across four lab-report extraction datasets.

Prior Google medical models followed similar patterns: Med-PaLM (arXiv:2212.13131) and Med-Gemini (arXiv:2404.18419) also reported benchmark improvements on MedQA and imaging tasks yet saw limited translation to deployed clinical systems, primarily because evaluations relied on retrospective internal datasets without prospective randomized trials or multi-site external validation. MedGemma 1.5 repeats this template; its reported metrics derive from curated benchmarks that do not replicate real-time hospital workflow conditions or demographic diversity required by FDA AI/ML device guidance.

Mainstream coverage emphasized benchmark numbers while missing the absence of reported radiologist-AI agreement studies, adverse-event tracking, or integration data with existing EHR vendors. The open release supplies a reproducible foundation, yet the report supplies no evidence on inference latency in clinical hardware, fine-tuning stability across institutions, or bias audits across global populations, leaving downstream developers responsible for the rigorous validation steps that determine actual patient impact.