Compute only where the signal is: Conformal Adaptation certifies sparse inference, holding Dice close on brain-tumor segmentation
Synopsis
Medical and astronomical images are orders of magnitude larger than natural images, yet tumors and galaxies occupy a tiny fraction of the data, and uniform compression such as quantization or pruning can erode precision exactly there. Conformal Adaptation first screens candidates with a low-resolution model, calibrates the selection threshold with conformal risk control so the discarded cells carry at most a user-set expected fraction of true signal mass, then refines only the selected regions. Across brain-tumor segmentation, head-and-neck segmentation, calcium spike-rate inference and galaxy flux super-resolution, the pipeline cuts computation while keeping accuracy close, and VISTA3D reaches an 82.10x FLOPs speedup in CPU-only inference.
Interpretation
Dense prediction no longer runs a full-resolution model at every location; a low-resolution model produces a coarse prediction, and high-resolution refinement is invoked only on regions likely to contain signal. Compression methods such as quantization and pruning scale all regions uniformly, which on sparse scientific data can sacrifice precision in the few regions that matter, and their error is assessed only after the fact. On brain-tumor segmentation, U-Net, Swin-UNETR and VISTA3D all keep Dice close to the full-resolution model after coarse prediction plus refinement while reducing floating-point operations; in CPU-only inference VISTA3D reaches an 82.10x FLOPs speedup.
The selection threshold is not set by hand but calibrated by conformal risk control, so the true signal mass in discarded cells is controlled in expectation. Earlier input-adaptive methods rely on learned or heuristic selection rules and offer no guarantee on the discarded information. Calibration uses a calibration set exchangeable with the test data, with loss defined as the fraction of signal mass in unselected cells; in brain-tumor segmentation the missed-cell fraction on calibration and test sets stays near the target 0.1 level.
Refinement touches only the selected coarse cells, and unselected cells are filled by upsampling or a null value, so cost scales with the number of selected cells rather than full-resolution sites. A full-resolution model must run the complete network at every location, with cost rising linearly with data scale. On the galaxy flux-field task, the calibrated pipeline retains only a fraction of coarse cells on average, with an analytic FLOPs speedup over the full-resolution baseline, at the cost of higher per-pixel mean absolute error and higher total-flux relative error than the full-resolution baseline, consistent with the certified bound on discarded flux.
Perspective
The demonstrated scope is dense prediction on uniform grids: MSD brain-tumor MRI segmentation, HECKTOR 2026 head-and-neck PET/CT segmentation, CASCADE calcium-imaging spike-rate regression and galaxy flux-field super-resolution on GalSim simulations; the authors note the derivation is domain-agnostic and extends to non-uniform domains such as graphs whenever the refinement operators preserve the hierarchical partition.
The guarantee bounds the expected fraction of true signal in discarded cells, not the error of the refined prediction; it is marginal, so an individual scan may exceed the set level, and small lesions that are faint at coarse resolution may be missed; it also assumes exchangeability between calibration and test data and is sensitive to covariate shift across scanners or sites. On the head-and-neck task, coarse plus refinement loses accuracy on GTVn, which the authors attribute to small lesions missed by the coarse model, and they list per-lesion risk control as future work.