A medical image is never observed directly. It is inferred from incomplete measurements, interpreted under genuine ambiguity, compressed into a usable description, and finally explained in terms of underlying biology. This talk uses that chain as its backdrop and takes up three distinct estimation problems situated at different points along it, each filling a different gap that the others leave open. The common stance is that each problem is best posed in terms of a distribution or an estimator rather than a single picture. The first problem is reconstruction, an ill-posed inverse problem: the image has to be recovered from far fewer measurements than there are unknowns, so the solution is not unique. Estimators that learn the prior structure of plausible images from data, while staying consistent with the measurements actually acquired, then single out a high-quality reconstruction. The second is interpretation under expert disagreement, where a single segmentation is untenable and the natural target becomes a predictive distribution, learned by matching expert variability through statistical divergences and assessed through calibration. The third is the move from image to biology: unsupervised generative models place whole-organ images on a low-dimensional manifold, yielding compact and reproducible descriptors of anatomy and function. Treated as quantitative phenotypes, these descriptors are then linked to the genome through heritability estimation and genome-wide association in tens of thousands of individuals, recovering and extending known genetic architecture. The thread throughout is that careful probabilistic and mathematical modelling, rather than any single method, is what turns images into biological insight.