MO‐D‐141‐08: Multi‐Scale Analysis of Collagen Architecture for Classifying Tumor and Healthy Breast Tissue Images

Abstract

Purpose: To quantify differences in collagen microarchitecture between healthy human breast tissue and breast cancer images using extended field of view second harmonic generation (EFOV SHG) microscopy coupled with analysis by the curvelet transform. Methods: Two frozen samples of normal associated human breast tissue (NAT) and invasive ductal carcinoma (IDC) were obtained through the University of Wisconsin BioBank. Samples were thawed, fixed, embedded in agar, cut into 300 micron thick sections using a vibratome and imaged using a custom inverted SHG microscope designed to allow for EFOV imaging. Individual SHG Images were captured at 0.79 micron resolution and a FOV of approximately 200 microns. The FOV was extended by stage scanning and image stitching to produce submicron resolution fields of view up to 2 cm2. NAT and IDC tissues were imaged, stitched using the Grid/Collection Stitching ImageJ plugin, and analyzed using features derived from the curvelet transform including variance in orientation and prevalence at various scales. These novel features were tested using support vector machine (SVM) cross validation. Results: IDC tissue contained linearly aligned, crisscrossing collagen while NAT tissue contained wavy, dense collagen. The mean curvelet alignment value for the normal images was approximately 2X higher than that for IDC. In addition, normal images had approximately 20% fewer closely correlated curvelets compared to IDC. Three‐fold cross validation of a linear SVM classifier revealed average sensitivity and specificity of 0.994 and 0.969 respectively. Conclusion: These results demonstrate the excellent potential for collagen architecture quantification based on features derived from the curvelet transform. We have also shown the value of EFOV SHG imaging to capture both large and small scale features necessary for quantifying collagen morphology. Future work will improve EFOV SHG imaging throughput and test classification capabilities of the curvelet based feature set on a large sample population. NIH R01 grants CA114462 and CA136590 to P.J.K. and K.W.E., T32 CA009206 to J.S.B and the Morgridge Institute for Research