A deep architecture for visually analyze Pap cells

This work proposes a deep ANN architecture which accomplishes the reliable visual classification of abnormal Pap smear cell. The system is driven by independent agents where the first agent consists of a three layer ANN pretrained to closely track a reticle pattern. This net participates in a local close loop that oscillates and produces unique time-space versions of the visual data. This information is stabilized and sparsed in order to obtain compact data representations, with implicit space time content. The obtained representations are delivered to second level agents, formed by independent three layers ANNs dedicated to learning and recognition activities. To train the system a noise-balanced algorithm is employed, where the training set is composed by pap cells and white noise. This combination operating on finite databases and in a self controlled learning loop, auto develops enough cell recognition knowledge as to classify whole classes of Pap smear cells. The system has been tested in real time utilizing documented data bases.