Neural network classification of laser-induced 5-ALA-PpIX fluorescence spectra using adaptive principal component extraction

A novel method of feature extraction and classification of fluorescence spectra using neural network was developed in order to improve the diagnostic rate of earlier stage colonic carcinomas with laser-induced 5-ALA-PpIX fluorescence spectra. 150 min after trail intravenous injections of 5-ALA dose of 25mg/kg body weight (BW) to 40 rats, 504 fluorescence spectra excited with 370 nm Ti-Laser were collected in vivo, which included 183 normal, 69 dysplasia (DYS), 87 early cancer (EC) and 165 advanced cancer (AC). After preprocessing, 6 principal components were extracted using adaptive principal component extraction (APEX). With BP neural network trained with resilient back-propagation algorithm (RBPNN), all spectra were divided into two categories: normal or abnormal, which included DYS, EC and AC. The sensitivity and specificity were 96.57% and 95.08% respectively. The accuracy of discriminating DYS and EC and AC from normal tissue were 92.75% and 98.85% and 96.36% respectively. The result indicated that this method could effectively diagnose earlier stage colonic carcinomas