Characterization of Collagen Fiber Organization in Breast Cancer via Model-Free Multiscale pSHG Image Analysis

Alterations in collagen micro-architecture are hallmarks of tumor progression. Conventional polarization second-harmonic generation (pSHG) analyses rely on rigid symmetry assumptions that often fail in heterogeneous tissue microenvironments. We present a fully automated model-free, multiscale, computational framework designed for the unbiased quantification of complex collagen organization in breast cancer tissue.

Breast tumor and adjacent perilesional tissues were imaged using a custom pSHG microscope and analyzed at micro- and meso-scale levels. Collagen centerlines were extracted via U-Net-based segmentation to estimate fiber orientations, while global alignment was quantified using 2D-FFT angular spectra. Structural organization was characterized with model-free descriptors, including scalar and biaxial order parameters and semi-variogram-based spatial autocorrelation.

Across a limited proof-of-concept dataset ( $n=14$ FoVs), the proposed multiscale framework effectively discriminated tumor from adjacent perilesional collagen architecture. The trigonal polarization model yielded smoother, more robust orientation maps than the cylindrical approach, showing strong agreement with deep learning fiber centerline analysis ( $CCC=0.985$ ). Tumor regions exhibited significantly longer spatial coherence length (87.3 ± 22.1 µm vs. 44.2 ± 11.3 µm, P = 0.01), higher scalar order (0.81 ± 0.06 vs. 0.56 ± 0.05, P = 0.01), and elevated biaxial order (0.12 ± 0.02 vs. 0.06 ± 0.06, P < 0.05).

This seminal framework provides a robust and assumption-free methodology to extract quantitative collagen descriptors across spatial scales. By integrating deep learning, frequency-domain analysis, and spatial statistics, it captures both local and long-range organizational features, supporting collagen architecture as a potential quantitative multiscale biomarker of tumor-associated extracellular matrix remodeling.