Infrared microspectroscopy and artificial neural networks in the diagnosis of cervical cancer

Infrared spectra of 88 normal and 32 abnormal (mild to severe dysplasia) cervical smear samples were used as a databank to investigate the usefulness of artificial neural networks (ANN) in the diagnosis of cervical smears. The spectra were first reduced, using principal component analysis (PCA), to seven wavenumber components that are the major contributors to the variance. A number of different ANN architectures were investigated that could differentiate between normal and abnormal cervical smears. Although the ANNs were trained to differentiate only normal from abnormal smears, the results using an independent test data set indicated that within the abnormal category mild dysplasia could be distinguished from severe dysplasia. The results using this restricted data set indicate that neural networks coupled to infrared microspectroscopy could provide an alternative automated means of screening for cervical cancer.