Abstract 4167: Virtual inference of collagen architecture from H&E to characterize stromal fiber morphology and organization in colorectal cancer

Abstract Background: Collagen architecture is a key determinant of colorectal tumor progression, influencing immune exclusion, fibroblast differentiation, and tumor budding, and plays a central role in shaping the tumor microenvironment. Although image analysis algorithms exist to quantify collagen features such as fiber density, orientation, and organization, but typically require specialized stains like picrosirius red (PSR). Computational methods that infer collagen phenotypes directly from H&E slides could enable larger studies of spatial tumor matrix biology and support development of prognostic imaging biomarkers. Methods: In this pilot study, four colorectal cancer slides were stained with H&E and imaged, then restained with PSR, and re-imaged. Co-registered stromal patches yielded 40,740 training/validation and 4,528 real PSR and H&E test image patch pairs. Two generative image-to-image translation models, pix2pix and a latent diffusion model were trained to generate virtual PSR (vPSR) patches from the H&E patches. The real PSR test patches paired with the H&E test patches served as reference. The virtual pix2pix (pix2pix-vPSR), diffusion (diffusion-vPSR) and real PSR test patches were processed to obtain collagen fiber masks. Subsequently, 108 collagen fiber density and orientation features were extracted from each mask and correspondence of the features extracted from pix2pix-vPSR and diffusion-vPSR collagen masks were compared with those extracted from real PSR collagen mask using Spearman correlation coefficients. Results: Both models produced strong correlations with real PSR features, particularly for collagen area fraction (r > 0.89), high-density matrix fraction (r > 0.86), gap circle count (r > 0.86), fiber spine area (r > 0.83), coherency/anisotropy (r > 0.82), fractal dimension (r > 0.82), fiber angle (r > 0.80), and fiber length (r > 0.80). Diffusion-vPSR correlations ranked among the top 90% and exceeded pix2pix-vPSR correlations by an average Δr > 0.07, with the gains observed for fiber contour- and thickness-based features, including median length (Δr > 0.20), thickness variation (Δr = 0.19), and asymmetry indices (Δr > 0.16). Smaller improvements were observed for fiber entropy length, variability and for skew/kurtosis/gap statistics (Δr=0.07-0.11). Pix2pix correlations surpassed diffusion correlations in only 10% of features; primarily gap fiber density features (Δr ≤ 0.05). Discussion: This study demonstrates the feasibility of generating vPSR images from routine H&E slides to infer collagen architecture with high fidelity. Broader application and refinement across larger cohorts will clarify how collagen architecture shapes tumor behavior, recurrence risk, and patient outcomes, supporting further spatial biomarker development. Citation Format: Harsimran Kaur, Minh-Khang Le, Arkadiusz Gertych, Maha Guindi, Jun Gong, Alexandra Gangi, Kyle Coleman, Jane C. Figueiredo, Xiaoying Liu, Louis J. Vaickus, Keluo Yao, Ken S. Lau, Joshua Jay Levy. Virtual inference of collagen architecture from H&E to characterize stromal fiber morphology and organization in colorectal cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4167.