Ischemic stroke is a leading cause of death and disability, yet decades of animal modelling and clinical trials have failed to deliver effective neuroprotection. A central obstacle is that clinical MRI cannot resolve the cellular state of tissue: what is lost, what is salvageable, and how far injury extends. Yet MRI guides patient selection, treatment windows and trial endpoints. No human ground truth has linked imaging to cellular pathology across the whole brain. This is the first whole-brain, cell-resolution histological reconstruction of human ischemic stroke registered to ante-mortem clinical MRI. Case 1 had a 16-day course after middle cerebral artery occlusion. We acquired in-vivo and post-mortem MRI, and serially sectioned the fixed brain at 20 µm (11,009 sections). Sections were stained with Nissl, H&E and silver (myelin), and by immunohistochemistry for fibrinogen (blood–brain barrier), HIF-1α (hypoxia), APP (axonal injury), GAP-43 (axonal sprouting), GFAP (astrocytes), CD68 (microglia/macrophages) and CD34 (vasculature). Sections were then reconstructed in 3D and registered to the in-vivo MRI. Case 2 (3-day course, posterior cerebral artery occlusion; 9,130 sections) provided qualitative histological corroboration. AI-based cell and pathology detection ran across the ~1.5-petabyte dataset. This enabled quantification by MRI-defined lesion, arterial territory and neuroanatomical region. Clinical MRI underestimated injury: histologically defined injury exceeded the DWI lesion by 39%. APP-labelled axonal injury and GAP-43-labelled sprouting extended through subcortical white matter beyond the infarct. They reached peri-infarct and distant tissue up to 40 mm from the core edge, including a potentially novel population of regenerating axons in the peri-infarct zone. Blood–brain barrier disruption was confined to the acute infarct and was negligible in distant and contralateral tissue. HIF-1α revealed hypoxic changes in distant cortex not evident on MRI, and subclinical pathology was present in the contralateral hemisphere. Case 2 showed concordant qualitative features. Cellular injury in human stroke is more extensive, distributed and dynamic than clinical MRI indicates. Hypoxic, inflammatory and reparative processes extend well beyond the imaging-defined penumbra. These findings redefine the target tissue and potential therapeutic windows, and expose the limits of MRI-based trial endpoints. They also provide a human-tissue benchmark for candidate mechanisms, including barrier protection, hypoxia signalling, neuroinflammation and axonal repair. This single-donor atlas is a proof of concept. The necessary next step is a multi-donor imaging–histology atlas spanning stroke subtypes, territories and time from onset. It should be large enough to train foundation models that infer cellular pathology from routine clinical MRI, making every scan a virtual biopsy. We call for a coordinated multi-centre brain donation and imaging programme to build it. An annotated subset of ~1,100 MRI-registered sections is openly available on this website; the full study is described in the accompanying paper: [link].