Cast aluminium alloys, and more widely cast materials, are frequently used in industry. The casting process allows for complex geometries of parts, but, on the downside, often causes materials voids. It is well known these material defects are harmful for material fatigue performances, but the nature of these defects, in a statistical manner, are more seldom studied. This paper aims at proposing a methodology for finding the underlying characteristics of the defect population (size and spatial distribution) and determine their implication on fatigue behaviour in the presence of stress/strain gradients (notched specimens). To do so, various statistical tools are brought from different fields, such as point processes, and applied to experimentally observed defect distributions (by μCT tomography on virgin test specimens). The population of defects is clearly identified, and it is shown these defects are not randomly distributed, but rather in cluster. It is also shown there is no strong link between the defect size an it’s location. Knowing the statistics of the defect population, it is then possible to confront the result of fatigue tests (and the observed initiating defects) with the simulated defect population: the fatigue crack initiation mechanisms, which favour (sub-) surface rather than core initiating defects, reduce the size of the active zone and therefore artificially shift the defect size distribution (by reducing their number).