Image quality assessment of imputation-induced biomarker distortions in multiplex immunofluorescence

Stain imputation in multiplex immunofluorescence (mIF) aims to reconstruct missing biomarkers but can inadvertently distort inter-marker dependencies essential for biological interpretation, leading to misleading downstream analyses. Conventional metrics such as Structural Similarity Index Measure or Mean Absolute Error quantify pixel-level fidelity but do not ensure the preservation of biologically meaningful relationships. We present a biologically informed image quality assessment (IQA) framework that assesses imputation-induced distortions in inter-marker relationships. Images with imputed CD8 generated by OS2CR-Diff and Stain Imputation in Multiplex Immunofluorescence Imaging (SIMIF) are labelled according to changes in CD8-CD16 colocalisation (inflated, deflated, or preserved states), and a deep learning model is trained to identify these states directly from the imputed images, enabling quality control of stain-imputation outputs. To address class imbalance in the curated dataset, we introduce a Correlation-Aware Augmentation (CAA) approach that perturbs CD8 spatial distributions relative to CD16, simulating imputation-relevant inflation and deflation patterns. CAA consistently improved classification performance (AUC and macro-F1) across both datasets, outperforming conventional augmentation approaches.