Second harmonic generation microscopy analysis of extracellular matrix changes in human idiopathic pulmonary fibrosis.

Patients with idiopathic fibrosis (IPF) have poor long-term survival as there are limited diagnostic/prognostic tools or successful therapies. Remodeling of the extracellular matrix (ECM) has been implicated in IPF progression; however, the structural consequences on the collagen architecture have not received considerable attention. Here, we demonstrate that second harmonic generation (SHG) and multiphoton fluorescence microscopy can quantitatively differentiate normal and IPF human tissues. For SHG analysis, we developed a classifier based on wavelet transforms, principle component analysis, and a K-nearest-neighbor algorithm to classify the specific alterations of the collagen structure observed in IPF tissues. The resulting ROC curves obtained by varying the numbers of principal components and nearest neighbors yielded accuracies of >95%. In contrast, simpler metrics based on SHG intensity and collagen coverage in the image provided little or no discrimination. We also characterized the change in the elastin/collagen balance by simultaneously measuring the elastin autofluorescence and SHG intensities and found that the IPF tissues were less elastic relative to collagen. This is consistent with known mechanical consequences of the disease. Understanding ECM remodeling in IPF via nonlinear optical microscopy may enhance our ability to differentiate patients with rapid and slow progression and, thus, provide better prognostic information.

Texture analysis applied to second harmonic generation image data for ovarian cancer classification.

Remodeling of the extracellular matrix has been implicated in ovarian cancer. To quantitate the remodeling, we implement a form of texture analysis to delineate the collagen fibrillar morphology observed in second harmonic generation microscopy images of human normal and high grade malignant ovarian tissues. In the learning stage, a dictionary of "textons"­frequently occurring texture features that are identified by measuring the image response to a filter bank of various shapes, sizes, and orientations­is created. By calculating a representative model based on the texton distribution for each tissue type using a training set of respective second harmonic generation images, we then perform classification between images of normal and high grade malignant ovarian tissues. By optimizing the number of textons and nearest neighbors, we achieved classification accuracy up to 97% based on the area under receiver operating characteristic curves (true positives versus false positives). The local analysis algorithm is a more general method to probe rapidly changing fibrillar morphologies than global analyses such as FFT. It is also more versatile than other texture approaches as the filter bank can be highly tailored to specific applications (e.g., different disease states) by creating customized libraries based on common image features.