DESIGN AND IMPLEMENTATION OF A MACHINE LEARNING-BASED EDGE-COMPUTING SYSTEM FOR BREAST CANCER SCREENING

Computation systems have been witnessing a paradigm shift in which processing tasks are migrating from resource-rich cloud servers to the resource limited edge nodes residing at the point where data are generated and possessed. This paradigm called as the edge-computing has become as the first option in the applications that data-privacy is of paramount importance such as medical diagnosis systems. We present the results of the feasibility study of a machine learning-based mammography interpretation tool that has been implemented on a low price commercial off the shelf edge computing hardware. We evaluated the proposed system using the models that reach area under the receiver operating characteristics (AUROC) of 0.857 and 0.881 for single view and four views interpretation, respectively. The evaluation results substantiate this configuration is able to rffectively run the machine learning models that performs breast cancer prediction using a single mammogram. Furthermore, we observe that computation resource demanded by more complex models that utilize four mammograms for the task of interpretation can be satisfied by the considered edge computing hardware.