Answers are grounded in this paper and cite their sources.
Layer-wise relevance propagation (LRP) is widely used to attribute the decisions of convolutional networks that screen Giemsa-stained blood films for malaria parasites, yet the stability of those attributions under routine stain-batch variation has not been characterised. the present analysis treats LRP attribution maps as estimators and asks how their agreement degrades when the input distribution shifts along the stain axis. We formalise attribution stability as a variance decomposition in which the between-batch component is the quantity of clinical interest, and we derive a propagation bound showing that the batch-induced perturbation of a relevance map is controlled by the product of the layer-wise gain of the network and the stain-induced input displacement. The bound implies that attribution instability is not a property of LRP alone but of the pairing between a network's local Lipschitz behaviour and the colour statistics of the staining protocol. We then specify a before-after-control-impact design for evaluating attribution stability across stain batches, with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), and report the operating characteristics of that design across 2000 replications per cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. Fifty-four clusters are required for 80% power at a 5-percentage-point effect. The contribution is a formal account of why LRP maps drift under stain variation and a design that makes that drift measurable before deployment.
No takes yet. Share an insight, caveat, or question.
Appadoo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: