The Anatomy Left Behind: Size Confounds in Neuron-Morphology Cell-Type Benchmarks

Neuronal cell types are defined by fine anatomical structure: the axonal cartridges of a chandelier cell, the ascending axon of a Martinotti cell. Self-supervised encoders of morphology are evaluated on how well they recover these labels, yet standard preprocessing removes the axon and subsamples the remaining tree to a few hundred points before training, so no encoder sees the structures the labels are defined by. We ask what these benchmarks measure. On the Allen Cell Types database a released encoder is outperformed by four features describing only a neuron's size and extent. Across six benchmarks spanning light and electron microscopy, most of each benchmark's above-chance performance is recoverable from those four features after its own preprocessing. The reasons differ by label: some labels are molecular or positional and never depended on shape; others are carried by branching structure that the preprocessing removes, and a pre-registered compartment ablation locates that structure in the axon on one corpus and, against prediction, in the dendrite on another. We propose no encoder. We recommend three one-line checks: report the size baseline, test which compartment carries the label, and evaluate preprocessed data against a model fitted to unreduced reconstructions.