Locally linear embedding method for dimensionality reduction of tissue sections of endometrial carcinoma by near infrared spectroscopy

Locally linear embedding (LLE) is introduced here as a nonlinear compression method for near infrared reflectance spectra of endometrial tissue sections. The LLE has been evaluated by using support vector machine (SVM) classifiers and the projected difference resolution (PDR) method. Synthetic data sets devised to resemble near-infrared spectra of tissue samples were used to characterize the performance of the LLE. The LLE was compared using principal component compression (PCC) method to evaluate nonlinear and linear compression. For a set of real tissue samples, if the compressed data were not range-scaled prior to SVM classification, the principal component compressed data gave an average prediction rate of 39 ± 2% while the LLE 94 ± 2%; if range-scaled after compression, the LLE and PCC performed evenly, with maximum average prediction values of 94 ± 2% and 93 ± 2%, respectively. The SVM without compression yielded a classification rate of 92 ± 2%. The prediction accuracy was consistent with PDR results. Without the second derivative preprocessing, the classification rates were 90 ± 3%, 89 ± 2%, and 78 ± 2% for the LLE compressed, the PCC, and no compression classifications by the SVM, respectively.

Figure optionsDownload as PowerPoint slideHighlights
► Locally linear embedding (LLE) is introduced first time to the field of spectroscopy as a dimensionality reduction method for feature extraction of near infrared spectra from reflectance measurements of endometrial tissue sections.
► LLE was evaluated and compared with principal component compression (PCC) by using support vector machine (SVM) classifiers.
► The projected difference resolution (PDR) method was used to evaluate the LLE method.
► Some exemplary synthetic data sets were created and NIR spectral data of real tissue samples were collected to verify LLE coupled to SVM for classification.
► LLE combined with the SVM gave better predictions and can effectively extract more discriminating features compared to PCC without scaling.