Quantification, radiomics and artificial intelligence in infection imaging: Current status and future directions in nuclear medicine

Abstract

Nuclear medicine infection imaging has traditionally relied on semantic visual interpretation supported by simple semi-quantitative indices. While effective, this paradigm is limited by observer dependence, restricted sensitivity to subtle or diffuse disease, and difficulty in standardising interpretation across centres. Advances in quantitative imaging, radiomics and artificial intelligence (AI) are reshaping this landscape. These complementary domains, collectively conceptualised as computomics, extend infection imaging from qualitative pattern recognition toward objective, reproducible and data-driven characterisation of disease. Quantitative imaging converts tracer distribution into measurable biological metrics, ranging from simple region-of-interest count ratios to standardised uptake values and kinetic parameters. Radiomics builds on this foundation by extracting high-dimensional features describing intensity, shape, texture and spatial heterogeneity, revealing image information not appreciable to the human eye. AI, through machine learning and deep learning approaches, integrates quantitative and radiomic data with clinical variables to automate segmentation, enhance reconstruction, support classification, and enable predictive modelling. Together, these tools offer potential to improve differentiation of infection from sterile inflammation, quantify disease burden, monitor therapy response, and standardise interpretation in complex scenarios. Quantitative accuracy and radiomic stability remain highly dependent on acquisition, reconstruction and processing parameters. AI-driven image enhancement and denoising may improve visual appearance while altering voxel statistics, with downstream effects on quantitative metrics and texture features. Variability in feature definitions, segmentation methods and analysis pipelines further limits reproducibility. Consequently, harmonisation, standardisation, transparent validation and physics-informed AI models are essential.

Nuclear medicine infection imaging has traditionally relied on semantic visual interpretation supported by simple semi-quantitative indices. While effective, this paradigm is limited by observer dependence, restricted sensitivity to subtle or diffuse disease, and difficulty in standardising interpretation across centres. Advances in quantitative imaging, radiomics and artificial intelligence (AI) are reshaping this landscape. These complementary domains, collectively conceptualised as computomics, extend infection imaging from qualitative pattern recognition toward objective, reproducible and data-driven characterisation of disease. Quantitative imaging converts tracer distribution into measurable biological metrics, ranging from simple region-of-interest count ratios to standardised uptake values and kinetic parameters. Radiomics builds on this foundation by extracting high-dimensional features describing intensity, shape, texture and spatial heterogeneity, revealing image information not appreciable to the human eye. AI, through machine learning and deep learning approaches, integrates quantitative and radiomic data with clinical variables to automate segmentation, enhance reconstruction, support classification, and enable predictive modelling. Together, these tools offer potential to improve differentiation of infection from sterile inflammation, quantify disease burden, monitor therapy response, and standardise interpretation in complex scenarios. Quantitative accuracy and radiomic stability remain highly dependent on acquisition, reconstruction and processing parameters. AI-driven image enhancement and denoising may improve visual appearance while altering voxel statistics, with downstream effects on quantitative metrics and texture features. Variability in feature definitions, segmentation methods and analysis pipelines further limits reproducibility. Consequently, harmonisation, standardisation, transparent validation and physics-informed AI models are essential.