Automatic detection of microcalcifications in mammography using a neuromimetic system based on retina

The incidence of breast cancer in France is roughly 26,000 and the annual number of deaths is 11,000. The mammography is the choice examination for the early identification of the tumours in an asymptomatic population. This is a simple, reliable, inexpensive examination, allowing to identify a grave and frequent pathology, but that can be the object of an effective treatment if early detected. The recognition of the microcalcifications in the mammographies is the key for early detection of cancers. Automatic detection methods were already proposed, but they have a very weak specificity and a relatively low sensibility. Currently, the eye of the expert still remains the better judge. We propose a neuromimetic method to localize automatically the microcalcifications. In this method, we devise a network of formal neurones inspired from the mammal retina architecture. This model mimics one characteristic of the retina which is is a sensor that automatically adapts to the image characteristics to analyse and realize the outlines extraction and adaptative filtering of the pictures, based on its network properties. The results were tested using a public standardized data set (DDSM), which was designed to test the automatic detection methods. We show that our "retina" can extracts most of the microcalcifications that can be grouped together in clusters. While we achieve a 95% sensitivity, we must acknowledge a low specificity (22%). Current efforts will focus to enhance this latter parameter.