Diagnosis of human coronary conditions by a neural network, with evolutionary wavelength selection in their quantised Raman spectra

A promising approach to in vivo detection of cardiovascular diseases in human beings would use a laser-based system to excite the inner walls of a patient's coronary, and an automatic system to analyse the resulting Raman radiation collected by an attached spectrometer. Here, an algorithm to perform such an automatic analysis is presented, which diagnoses three conditions of human coronaries - normal, atheromatous and calcified. The algorithm implements three general ideas: frequency selection via the genetic algorithm, for achieving compression and redundancy elimination in the data; neural network learning, which is meant to embody the actual solver of the inverse problem at issue, namely, the diagnosis of the coronary condition, given its Raman spectrum; and quantisation of the collected spectra, for speeding up the neural network learning and helping to reduce redundancy in the intensity levels. The resulting method is conceptually simple and achieves very good performance, while still being amenable to further natural improvements, of significant potential impact. In order to support this statement, experiments are discussed on the diagnosis of coronary conditions of real subjects, from very noisy Raman spectra, and a comparison is made between our results with those obtained from two competing approaches.