Biophotonics methods are attractive since they allow for the non-invasive diagnosis of cancer. Experiments
were carried out to investigate the feasibility of detecting early pre-cancer using optical spectroscopy. However,
optimization of instrumentation design parameters remains challenging because of the lack of metrics to evaluate
the performance of certain design parameters. For example, although using angled-collection geometry has been
shown to collect depth sensitive spatial origins, the performance of devices with angled-collection geometries
are not well characterized or quantified. In this study, we use a polarization-sensitive Monte Carlo simulation
(Pol-MC) to aid in the design of instrumentation for the early detection of epithelial cancer. The tissue is
modeled in layers: (0) air outside the tissue, (1) epithelial layer, (2) thin pre-cancer layer of cells, (3) thin
basement membrane, implemented as a thin transparent layer, and (4) the stroma, implemented as a thick layer
of scattering material. We propose a new metric, Target Signal Ratio (TSR), to evaluate the proportion of signal
that is scattered from a target layer, which is the basal/pre-cancer layer. This study is a proof-of-concept for
the application of computational techniques to facilitate instrument design.
We present an approach to adaptively adjust the spectral window sizes for optical spectra feature extraction. Previous studies extracted features from spectral windows of a fixed width. In our algorithm, piecewise linear regression is used to adaptively adjust the window sizes to find the maximum window size with reasonable linear fit with the spectrum. This adaptive windowing technique ensures the signal linearity in defined windows; hence, the adaptive windowing technique retains more diagnostic information while using fewer windows. This method was tested on a data set of diffuse reflectance spectra of oral mucosa lesions. Eight features were extracted from each window. We performed classifications using linear discriminant analysis with cross-validation. Using windowing techniques results in better classification performance than not using windowing. The area under the receiver-operating-characteristics curve for windowing techniques was greater than a nonwindowing technique for both normal versus mild dysplasia (MD) plus severe high-grade dysplasia or carcinama (SD) (MD+SD) and benign versus MD+SD. Although adaptive and fixed-size windowing perform similarly, adaptive windowing utilizes significantly fewer windows than fixed-size windows (number of windows per spectrum: 8 versus 16). Because adaptive windows retain most diagnostic information while reducing the number of windows needed for feature extraction, our results suggest that it isolates unique diagnostic features in optical spectra.
We report the results of an oral cavity pilot clinical trial to detect early precancer and cancer using a fiber optic probe with obliquely oriented collection fibers that preferentially probe local tissue morphology and heterogeneity using oblique polarized reflectance spectroscopy (OPRS). We extract epithelial cell nuclear sizes and 10 spectral features. These features are analyzed independently and in combination to assess the best metrics for separation of diagnostic classes. Without stratifying the data according to anatomical location or level of keratinization, OPRS is found to be sensitive to four diagnostic categories: normal, benign, mild dysplasia, high-grade dysplasia, and carcinoma. Using linear discriminant analysis, separation of normal from high-grade dysplasia and carcinoma yield a sensitivity and specificity of 90 and 86%, respectively. Discrimination of morphologically similar lesions such as normal from mild dysplasia is achieved with a sensitivity of 75% and specificity of 73%. Separation of visually indistinguishable benign lesions from high-grade dysplasia and carcinoma is achieved with good sensitivity (100%) and specificity (85%), while separation of benign from mild dysplasia gives a sensitivity of 92% and a specificity of 69%. These promising results suggest that OPRS has the potential to aid screening and diagnosis of oral precancer and cancer.
We propose an approach to adaptively adjust the spectral window size used to extract features from optical
spectra. Previous studies have employed spectral features extracted by dividing the spectra into several spectral
windows of a fixed width. However, the choice of spectral window size was arbitrary. We hypothesize that
by adaptively adjusting the spectral window sizes, the trends in the data will be captured more accurately.
Our method was tested on a diffuse reflectance spectroscopy dataset obtained in a study of oblique polarization
reflectance spectroscopy of oral mucosa lesions. The diagnostic task is to classify lesions into one of four
histopathology groups: normal, benign, mild dysplasia, or severe dysplasia (including carcinoma). Nine features
were extracted from each of the spectral windows. We computed the area (AUC) under Receiver Operating
Characteristic curve to select the most discriminatory wavelength intervals. We performed pairwise classifications
using Linear Discriminant Analysis (LDA) with leave-one-out cross validation. The results showed that for
discriminating benign lesions from mild or severe dysplasia, the adaptive spectral window size features achieved
AUC of 0.84, while a fixed spectral window size of 20 nm had AUC of 0.71, and an AUC of 0.64 is achieved
with a large window size containing all wavelengths. The AUCs of all feature combinations were also calculated.
These results suggest that the new adaptive spectral window size method effectively extracts features that enable
accurate classification of oral mucosa lesions.
Proc. SPIE. 6515, Medical Imaging 2007: Image Perception, Observer Performance, and Technology Assessment
KEYWORDS: Computer aided diagnosis and therapy, Data modeling, Calibration, Image processing, Computing systems, Receivers, Medical imaging, Artificial neural networks, Image enhancement, Probability theory
In this study we developed an effective novel method for reducing the variability in the output of different artificial neural network (ANN) configurations that have the same overall performance as measured by the area under their receiver operating characteristic (ROC) curves. This variability can lead to inaccuracies in the interpretation of results when the outputs are employed as classification predictors. We extended a method previously proposed to reduce the variability in the performance of a classifier with data sets from different institutions to the outputs of ANN configurations. Our approach is based on histogram shaping of the outputs of all ANN configurations to resemble the output histogram of a baseline ANN configuration. We tested the effectiveness of the technique using synthetic data generated from two two-dimensional isotropic Gaussian distributions and 100 ANN configurations. The proposed output calibration technique significantly reduced the median standard deviation of the ANN outputs from 0.010 before calibration to 0.006 after calibration. The standard deviation of the sensitivity of the 100 ANN configurations at the same decision threshold reduced significantly from 0.005 before calibration to 0.003 after calibration. Similarly the standard deviation of their specificity values decreased significantly from 0.016 before calibration to 0.006 after calibration.