Evaluation of Raman spectroscopic mapping for the detection of cervical intraepithelial neoplasia

Objektive: Cervical cancer is the second most frequent cancer among women worldwide. For the early detection of precancerous tissue, i.e., cervical intraepithelial neoplasia (CIN), lposcopically-guided biopsy and a following histological analysis is currently the gold standard with fairly acceptable sensitivity and specificity. However, colposcopy requires a physician to be highly experienced and histopathological results can be conflicting due to high levels of inter- and intra-observer discrepancies. The development of a method for the non-invasive, objective detection of precancerous lesions in real-time has been the goal of many studies using optical spectroscopy so far. Previous Ramanspectroscopic studies of cervical precancer used fiber probes or microspectroscopy and have reached sensitivities and specificities in the range of 70–100%, while influences of hormonal status, previous disease and inflammation on the spectra has been reported.

The goal of our current prospective study is to evaluate the robustness of our Ramanspectroscopic method, which employs macroscopic Raman mapping of large cervical samples.

Material and methods: Macroscopic mapping of the sample surface was performed in order to obtain data from more than one CIN stage and from regions of normal tissue from the one patient.

For this study, a customized Ramanspectroscopy system was built based on commercial components including a 785nm laser for excitation.

The cervical tissue samples were obtained by loop excision (conisation) following the clinical routine.

An overview color image of the sample including the grid of measurements points was acquired prior to the measurement. After the Raman measurement (in saline solution), the samples were fixed and sent to the pathologist, who documented the results of the histopathological analysis on the image of the sample from before the Raman measurement. Only spectra from within a homogeneous region of a certain tissue type were used as input data for the discriminant analysis. A reduced set of spectral features gained from principal-component analysis or wavelet transformation was used for a k-nearest neighbour discriminant analysis. A leave-one-patient-out training was used for estimating the diagnostic performance.

Results: A preliminary discriminant analysis of the data from the first 6 patients which have been measured under optimized experimental settings results in a sensitivity and specificity of 92% and 85% for the discrimination of CIN2–3 (65 spectra) from normal squamous epithelium (120 spectra).

Conclusion and Outlook: The discrimination of high grade lesions (CIN2–3) from normal squamous epithelium seems possible. The influence of the method of discriminant analysis and the number of available spectra and samples on the performance as well as the discrimination of CIN2–3 from low grade changes (CIN1) and other types of normal tissue (metaplasia, columnar epithelium) will be investigated.

Spatially resolved visualization of the classification results will give further insides on the performance of the method.

Acknowledgement: This project is co-financed by the Senate of Berlin and by the European Union (European funds for regional development, FKZ 10147189).