Stain Imputation in Multiplex Immunofluorescence Imaging (SIMIF) Based on Random Channel-Wise Masking

Recent advances in digital imaging have fuelled interest in using multiplex immunofluorescence (mIF) images to study multiple biomarkers and their interactions within a single tissue of the tumour microenvironment. However, the mIF data remain less accessible due to the need for specialised equipment and costly reagents, which increases the technical complexity, expense and time required. Additionally, issues like misalignment and artefacts can result in unusable or incomplete data, highlighting the need for effective stain imputation methods. The current state-of-the-art (SOTA) stain imputation method relies on supervised generative deep-learning models with a fixed panel of biomarkers, which often fail when some of the required biomarkers are absent. To address this limitation, we propose a novel stain imputation method for mIF (SIMIF), which integrates the Wasserstein Generative Adversarial Network (GAN) with a random channel-wise masking (RCWM) training strategy. This method effectively enhances robustness in handling various input biomarkers, leading to a more stable optimisation process and improved synthetic image quality. Evaluated on a dataset of 33,265 image patches extracted from 36 mIF whole-slide images, SIMIF demonstrates clear superiority over the current SOTA, achieving marked improvements in imputation performance for both CD8 and PD-L1 biomarkers.