Priors in Medical Image Analysis: A Review

Abstract

Medical imaging tasks are frequently limited by data that is either insufficient or excessive. In tomographic reconstruction, interventional fluoroscopy and endoscopic imaging, the available measurements are incomplete: Projections are few, the angular range is narrow, only one or two views are acquired or the observations are corrupted by physical effects that the imaging model does not account for. In digital pathology the situation is reversed, since a single whole-slide image may contain 109-1010 pixels and 104-105 usable tiles and therefore cannot be processed at native resolution. This review examines both situations and notes that each is addressed by the same fundamental mechanism, namely the introduction of a prior that constrains the space of admissible solutions. Over the past three decades the necessity of the prior has not changed, but its form has: Analytic smoothness penalties were succeeded by hand-designed variational regularizers and patch-based self-similarity models and subsequently by learned regularizers embedded in unrolled iterations, untrained network architectures, generative score models, attention distributions over gigapixel fields and large-scale foundation encoders whose embeddings provide task-agnostic representations. We review this development across five application areas image restoration, sparse-view and limited-angle tomography, biplane interventional imaging, endoscopic vision and computational pathology and summarise the limitations that remain unresolved, including hallucination of structure in the undetermined null space, poorly calibrated uncertainty, covariate shift across scanners and institutions and evaluation metrics that correlate weakly with diagnostic performance.