The problem — A segmentation or classification model trained on fluorescence data may assign 94% confidence to a prediction it gets wrong. Without calibration, downstream logic that thresholds on confidence — flagging rare phenotypes, routing to manual review, gating a readout — operates on numbers that carry no probabilistic meaning. In high-content imaging pipelines, overconfident wrong calls are silent: the pipeline completes, the numbers look clean, and the biology is wrong.
What it is / how it works — Temperature scaling divides every output logit by a single scalar T before the softmax. When T > 1 the distribution softens — the model becomes less certain — and when T < 1 it sharpens. T is fit on a held-out validation set by minimising negative log-likelihood; it costs one extra parameter and does not change which class wins, only how strongly the model asserts it. Guo et al. (ICML 2017) showed that this trivially simple transform reliably outperforms more elaborate post-hoc methods on modern deep networks.
Where it breaks — Temperature scaling fits one global T: it cannot correct miscalibration that varies by cell type, staining batch, or image quality tier. A single plate with unusual background fluorescence may arrive well-segmented but still confidently mislabelled at the readout stage. Calibration should therefore be validated per-condition rather than assumed to transfer across experimental contexts. See uncertainty-drift-and-failure-modes for the broader failure landscape, and validation-without-ground-truth for strategies when a labelled calibration set is unavailable.
The Fovea take — We treat calibration as a pipeline stage, not a model property. Before any confidence score gates a biological conclusion, it is checked against a held-out empirical frequency on data that matches the deployment distribution. Temperature scaling is the default first pass: cheap, interpretable, and well-understood. When batch or domain shift is present, per-stratum recalibration is added. The goal is outputs that mean what they say — a prerequisite for the verified, reproducible pipelines that distinguish production-grade computational microscopy from proof-of-concept work.