88 Machine learning-based collagen fiber quantification enables analysis of the pancreatic cancer tumor microenvironment directly from hematoxylin and eosin-stained whole slide images
Year: 2024
Type: conference-abstract
Abstract: Background Pancreatic ductal adenocarcinoma (PAAD) is notoriously resistant to immunotherapy. The desmoplastic stroma of PAAD may act as a barrier against T-cell infiltration. However, the relationship between collagen and anti-tumor immunity is complex, as stromal collagen has been proposed to constrain PAAD progression. To investigate the implications of collagen in PAAD, we developed a machine learning (ML) method to extract and quantify collagen features directly from hematoxylin and eosin (H&E)-stained whole slide images (WSI). Methods Quantitative Multimodal Anisotropy Imaging (QMAI), a polarization-based method to image tissue components with birefringence (e.g., collagen fibers), was performed on tissue sections (n=252) from 13 cancer types (figure 1A). Brightfield H&E WSI were used to match polarization WSI at pixel-level accuracy (figure 1B). These images were used to train a model (inferred QMAI, or iQMAI) to perform virtual polarization imaging on digital H&E images, allowing features of collagen to be extracted and computed (figure 1C). iQMAI was deployed on H&E-stained PAAD WSI (TCGA; n=204), and fibers were extracted as described.1 Cancer stroma was identified using PathExploreTM tissue models*. The LRRC15-CAF signature2 was calculated in cases with transcriptomic data (n=128). Association between collagen density and overall survival (OS; n=132) was assessed using Cox proportional hazard modeling, with disease stage as a covariate. *For research use only. Not for use in diagnostic procedures. Results Quantitative assessment of collagen features in PAAD using iQMAI revealed high variability in collagen density, with ~4-fold variation across the cohort. Stromal fiber density correlated negatively with LRRC15-CAF gene signature (associated with poor immunotherapy responses2) (Spearman rho=-0.22, p=0.014). Finally, high stromal collagen density was associated with longer OS (HR=0.65 (0.47, 0.90), p=0.009). Conclusions iQMAI, a physics-based ML technique trained on multiple cancer types, allows measurement of collagen directly from H&E WSI, enabling investigation of the interplay between collagen and other TME components. We identify a complex relationship between collagen, CAF subtypes, and survival in PAAD. Gene expression consistent with the immunosuppressive LRRC15 CAFs was associated with less abundant collagen, while high fiber density was associated with improved OS. Therefore, abundant stromal collagen may be protective in PAAD, consistent with previous immunohistochemistry-based results. Understanding the relationship between collagen, lymphocyte infiltration, and disease progression may aid the development of effective immunotherapies in PAAD. References Bredfeldt JS, Liu Y, Pehlke CA, Conklin MW, Szulczewski JM, Inman DR, Keely PJ, Nowak RD, Mackie TR, Eliceiri KW. Computational segmentation of collagen fibers from second-harmonic generation images of breast cancer. J Biomed Opt 2014;19:016007. Dominguez CX, Müller S, Keerthivasan S, Koeppen H, Hung J, Gierke S, Breart B, Foreman O, Bainbridge TW, Castiglioni A, Senbabaoglu Y, Modrusan Z, Liang Y, Junttila MR, Klijn C, Bourgon R, Turley SJ. Single-cell RNA sequencing reveals stromal evolution into LRRC15+ myofibroblasts as a determinant of patient response to cancer immunotherapy. Cancer Discov 2020;10:232-253.
Institution: PathAI (United States)
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4404065295