The number seems to come from the work of Edelstyn et al. They used a lumbar spine specimen obtained from autopsy of a 57-year-old woman and drilled conical holes of various sizes into the cancellous bone to simulate metastatic disease. They placed the spine in a water phantom to represent the abdomen, and ensured that the holes filled with water to represent tumor tissue.
They found that somewhere between 50%-75% loss of bone thickness was needed before the lesion could be detected on a lateral radiograph of the spine. That is, at 50% bone loss, 4 radiologists were unable to detect the lesion, but at 75% (the next grade up), all 4 were able to detect the lesion. The lesions were harder to detect on the frontal view, where between 67% and 100% loss of bone thickness was needed for identification.
They noted that these findings would be easier to detect in younger patients, where bone mineral density is preserved, and harder to detect in older patients.
They then drilled a series of small holes in the cortex and noted that detection was easier when the defect was tangential to the beam.
In summary, the numbers apply to lesions in the cancellous bone of a disembodied lumbar spine of a middle-aged woman placed in a water phantom as detected on a lateral radiograph. They do not apply to the frontal view of the spine, to cortical bone, or to cancellous bone elsewhere in the body.