Treating electrode coordinates as learnable parameters: percept-aware optimization on folded cortex improves reconstruction fidelity while eliminating vascular safety-margin violations

A percept-aware optimization framework for three-dimensional cortical electrode placement that directly minimizes predicted task-relevant perceptual error rather than geometric coverage metrics. Prior work either optimized stimulation parameters for fixed implants or arranged electrodes to maximize visual field coverage; this work makes electrode coordinates themselves learnable parameters optimized end-to-end within a differentiable prosthetic vision forward model. On the FreeSurfer fsaverage template, across electrode counts N∈{64–1024} and phosphene spreads ρ∈{500–1500} µm with 3 random initializations per configuration, compared against visual field tiling and visual field coverage baselines using Wilcoxon signed-rank tests at p≤0.01; on MNIST median MSE fell 67.7% versus tiling and 33.4% versus coverage, SSIM rose 11.1% and 4.7%, and downstream classification accuracy rose 62.6% and 22.4%.

Vascular avoidance and gray matter feasibility are explicitly integrated as differentiable penalty terms in the placement optimization, so functional optimization and surgical safety are addressed jointly in one framework. Coverage-driven placement strategies were safety-aware but did not jointly solve safety constraints and perceptual objectives in a single optimization; this work uses a 300 µm hinge-style vascular penalty and a gray matter signed-distance penalty to allow continuous trade-offs. Without safety constraints a large fraction of electrodes violated the 300 µm margin; incorporating the vascular penalty eliminated all margin violations while SSIM decreased by only 1.7% (MNIST) and 4.4% (CIFAR-10) relative to unconstrained optimization.

The framework extends to multi-electrode threads, jointly optimizing entry location and insertion trajectory under a fixed number of cortical insertions and improving perceptual fidelity. Device architecture (threads) and surgical constraints are brought into the same percept-aware objective, enabling quantitative exploration of device design trade-offs rather than only single-electrode position optimization. Under a fixed insertion budget of Ninsert = 128, threaded configurations improved perceptual fidelity relative to single-electrode placements (Fig. 4); optimizing a single configuration required less than 10 minutes on an NVIDIA RTX 3090.

The framework targets pre-operative planning, suited to researchers and device designers who determine electrode coordinates within 3D cortical gray matter and weigh perceptual performance against vascular safety under a given device electrode count; results are reported in the unified anatomical reference space of the FreeSurfer fsaverage template, and the authors state the framework is fully compatible with patient-specific cortical reconstructions, retinotopic mapping (e.g., fMRI), and angiographic data, so its intended setting is a planning pipeline that can be rerun on individual anatomy. The threaded extension addresses joint optimization of entry location and trajectory under a fixed insertion budget (Ninsert = 128), enabling quantitative exploration of device design trade-offs.

The perceptual forward model uses a simplified Gaussian phosphene with linear superposition, and the authors state future work will incorporate more detailed biophysical models; experiments were conducted on the fsaverage template rather than subject-specific anatomy, leaving performance under patient-specific reconstructions, fMRI retinotopy, and angiographic data an open question; gray matter confinement serves as a proxy for targeting layer 4, with laminar boundaries not explicitly modeled; evaluation relies on simulated MNIST and CIFAR-10 tasks with a ResNet-50 classifier and does not involve in vivo or clinical outcomes; threaded results are reported under a fixed insertion budget, and trade-offs under other budgets and device parameters remain to be explored.