Clinical Computer Vision for Cervical Cytology Under Stain Variation Across Low-Resource Laboratories: A Domain-Adaptation Evaluation of Cancer Screening Sensitivity Loss

Stain variation across low-resource cytology laboratories is widely acknowledged as a threat to the sensitivity of computer-vision cervical screening, yet the size of the sensitivity loss and the conditions under which domain adaptation recovers it remain poorly quantified. The framework decomposes stain-induced distribution shift into a linear colour-operator component and a nonlinear chromatin-texture component, and shows that the two components propagate differently through a convolutional classifier: the linear component is largely removable by colour normalisation, whereas the texture component produces a sensitivity loss that normalisation cannot address. A design analysis was conducted under a declared variance structure with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.64, across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The analysis indicates that 54 clusters are required to detect a 5-percentage-point adaptation effect at 80% power. The dominant obstacle to reliable adaptation evaluation is not the adaptation algorithm but the cluster-level heterogeneity of stain protocols, and that sensitivity-loss claims should be reported with the design parameters that produced them.