In this study, we proposed that retromammary space of a mammogram can be automatically detected, and breast positioning was classified based on quality control and validation using DCNNs. For the automatic classification, the accuracy was approx. 58%, the highest value obtained using VGG16. Although the initial results were low, the study demonstrated that evaluating positioning techniques in mammography using automation is feasible. Improving accuracy requires adjusting the learning image localization. Further improvement in accuracy is expected by performing other various image density corrections.

The Softmax values are a commonly used activation function in DCNN for image classification. The output of the fully connected layer is finally fed to the softmax function. The fully connected layer output is fed into the softmax function for final classification, which ensures that our predictions add up to 1. Although the conventional visual classification in guidelines evaluated only scale evaluation (3- or 4-scale), the softmax function could obtain more detailed evaluation metrics. Radiological technologists can use these metrics to assess and enhance their imaging techniques.

This study suggested the feasibility of automated evaluation of mammographic retromammary space using DCNN. The probability output from softmax can serve as a quantitative indicator of the imaging technique. By checking the index of the captured image, it could consider positioning. It is expected that feedback from AI will help radiographers improve their skills.