Threshold transfer for CoRegOVCD’s final mask inference

Determine whether and under what conditions the fixed decision thresholds used in CoRegOVCD’s final mask inference can transfer across datasets, classes, and imaging conditions without dataset-specific tuning, and establish principled strategies for threshold selection if transferability is limited.

Background

CoRegOVCD converts a pooled score map into a binary mask using an 8-bit quantization and a fixed threshold, with thresholds and related decoding hyperparameters fixed per evaluation configuration. While this yields strong results, it remains unclear how well such thresholds generalize across datasets and classes without retuning, prompting the open question of threshold transferability.