A validation model for segmentation algorithms of digital mammography images

We present a comprehensive validation analysis to evaluate the performance of three existing digital mammography segmentation algorithms against manual segmentation results produced by two expert radiologists. This is an improvement of an early methodology used for the evaluation of boundary algorithms on medical images. The mammography images used were acquired from the Digital Database for Screening Mammography (DDSM) and subsequently used as ground truth. We assessed threeexisting segmentation algorithms: (a) the Region Growing Combined with the Maximum Likelihood (ML) Model, (b) the Gradient Vector Flow (GVF) Model,and (c) the Standard Potential Field (STD) Model. We applied a comprehensive statistical metric. We concluded that the Region Growing Combined with the Maximum Likelihood (ML) Model yielded not only the best accuracy, specificity, percent error, and algorithm ranking, but also the greatest ratio of average computer-to-observer agreement and average inter-observer agreement (WI’). We also noted that the upper limit of the 95% Confidence Interval (CI) was greater than 1.0, and thus each individual observer is a reliable member of the group. These studies are especially important for the development of computer-aided diagnosis (CAD) systems for cancer.