Hollow axles, widely employed in High Speed and Very High Speed railways applications, are typically inspected by the ultrasonic testing method applied using a suitable boreprobe roto-translating along the longitudinal bore. Different ultrasonic transducers, emitting shear waves at 2-4 MHz and at different refraction angles in steel, are used to inspect the whole external surface of the axle in order to individuate cracks and defects at press-fit seats and at the body, especially along geometrical transitions. The structural integrity of the wheel-set is then based, together with other factors, on the efficient and reliable determination of the Probability of Detection (POD) curve of the boreprobe. For a given probes configuration, the POD curve is function of UT equipment, its calibration and its inspection parameters, therefore the same UT equipment could develop different POD curves as a function of the operating procedure normally defined by wheel-set manufacturers or end users. It is then important to split the absolute performances of the UT equipment by the different performances obtainable with different calibration and operation procedures. In order to derive such curves, a statistically representative sample of natural defects is needed. For this purpose, ten full-scale hollow axles, made of a quenched and tempered alloyed railway steel, were fatigue tested using a dedicated bench in order to develop two natural fatigue cracks in each of them. Different artificial defects were also introduced on the external surface on one of the axles by EDM. A total set of 28 semi-circular defects (17 natural and 11 artificial) was then obtained in a range of depths from 0.4 mm to 12 mm. Such a set allowed the experimental derivation of the POD curve of the boreprobe following the “signal response” approach. A general analysis of different calibration method leads to a comparison between different POD obtainable with the same UT inspecting device. A final part of the research consisted in the optimization of the boreprobe itself, by means of suitable numerical simulations, in terms of frequency, dimension of the transducer, angle of refraction and location of the defect/crack.