How Do Lesion Size and Random Noise Affect Detection Performance in Digital Mammography?

Abstract

We investigated the effect of random noise and lesion size on detection performance in mammography. Digital mammograms were obtained of an anthropomorphic breast phantom with and without simulated mass lesions. Digital versions of the mass lesions, ranging in size from 0.8 to 12 mm, were added back to the breast phantom image. Four alternate forced choice experiments were performed to determine the lesion contrast required to achieve a 92% correct lesion detection rate, denoted I92. Experiments were performed using identical phantom images and different versions of phantom images obtained using the same techniques but with different random noise patterns. For lesions larger than 1 mm, the slope of the contrast detail curves was always positive. This behavior contrasts with conventional contrast-detail curves in uniform backgrounds in which the slope is approximately -0.5. There was no difference between twinned experiments and those obtained using different patterns of random noise for lesions greater than 1 mm. When the lesion size was reduced below 1 mm, the detection threshold increased indicating a deterioration of lesion detectability, and detection performance was significantly lower when random noise patterns were used. Our results suggest that lesion detection is dominated by anatomical structure for lesions with a size >1 mm, but by random noise for submillimeter sized lesions.