Bringing Photorealistic Anatomy Beyond the Workstation with Gaussian Splatting: Opportunities, Limitations, and Validation Priorities

Three-dimensional Gaussian splatting is the latest computer-graphics technique to be adopted for medical visualisation, following volume rendering and cinematic rendering. Unlike its predecessors, which render images directly from the acquired volumetric data, Gaussian splatting is a reconstruction: a cloud of anisotropic Gaussian primitives is optimised against a finite set of rendered two-dimensional views, after which the source volume plays no further part. This distinction has practical consequences that are not evident from the exterior images typically used to demonstrate the technique. In this narrative review we describe what we term the view-space limitation: because training cameras orbit an anatomical structure from outside, the interior is never observed, and no clipping-plane position recovers intracardiac anatomy from a cardiac splat. We argue that this is a structural consequence of camera distribution rather than an implementation defect, that a splat is an inferred representation whose geometric and densitometric fidelity must be measured rather than assumed, and that until such validation is reported, splat reconstructions should be reserved for communication, education, and qualitative review rather than primary quantitative interpretation. We also argue that the case in favour is stronger than usually made: a splat of a patient’s heart renders at interactive frame rates in a consumer smartphone browser and can be shared by a link, extending photorealistic anatomy to referring clinicians, multidisciplinary meetings, trainees, patients, and health systems without workstation-class hardware. We propose a framework for matching visualisation technique to clinical setting and outline validation priorities, including accuracy against ground truth, projection-domain sparse-view formulations, clippable variants, and dynamic extensions for gated cardiac data, that should precede further application papers.