Computer-assisted quantification of tumor-associated collagen signatures to improve the prognosis prediction of breast cancer

Abstract

Background Collagen fibers play an important role in tumor initiation, progression, and invasion. Our previous research has already shown that large-scale tumor-associated collagen signatures (TACS) are powerful prognostic biomarkers independent of clinicopathological factors in invasive breast cancer. However, they are observed on a macroscale and are more suitable for identifying high-risk patients. It is necessary to investigate the effect of the corresponding microscopic features of TACS so as to more accurately and comprehensively predict the prognosis of breast cancer patients. Methods In this retrospective and multicenter study, we included 942 invasive breast cancer patients in both a training cohort (n = 355) and an internal validation cohort (n = 334) from one clinical center and in an external validation cohort (n = 253) from a different clinical center. TACS corresponding microscopic features (TCMFs) were firstly extracted from multiphoton images for each patient, and then least absolute shrinkage and selection operator (LASSO) regression was applied to select the most robust features to build a TCMF-score. Finally, the Cox proportional hazard regression analysis was used to evaluate the association of TCMF-score with disease-free survival (DFS). Results TCMF-score is significantly associated with DFS in univariate Cox proportional hazard regression analysis. After adjusting for clinical variables by multivariate Cox regression analysis, the TCMF-score remains an independent prognostic indicator. Remarkably, the TCMF model performs better than the clinical (CLI) model in the three cohorts and is particularly outstanding in the ER-positive and lower-risk subgroups. By contrast, the TACS model is more suitable for the ER-negative and higher-risk subgroups. When the TACS and TCMF are combined, they could complement each other and perform well in all patients. As expected, the full model (CLI+TCMF+TACS) achieves the best performance (AUC 0.905, [0.873–0.938]; 0.896, [0.860–0.931]; 0.882, [0.840–0.925] in the three cohorts). Conclusion These results demonstrate that the TCMF-score is an independent prognostic factor for breast cancer, and the increased prognostic performance (TCMF+TACS-score) may help us develop more appropriate treatment protocols.

A deep learning-based stripe self-correction method for stitched microscopic images

“…7. Fourth, tumor-associated collagen signatures are associated with the tumor development and disease prognosis 10,28 . In the Fig.…”

“…However, the stripes, shadings, and even artifacts often remain in the stitched fluorescence images, especially in the weak signal regions of label-free or large-scale stitched images [5][6][7][8][9] . According to our imaging experience and previous research results, it has been discovered that, even for the most stable commercial nonlinear optical microscopic instruments, the diverse stripes may still exist in the acquired images 10 . Optical engineers often underestimate the impact of stripes on biomedical researches.…”

“…However, many latent factors, e.g., laser source, detector noise, specimen quality, and stability of optical devices, jointly affect the shading patterns of image tiles during practical image acquisition. As a result, the multi-factor influenced shading patterns appear to be inconsistent, or even unpredictable, in a large-scale stitched image, thus appear various types of stripes or artifacts [5][6][7]10,12,13,[32][33][34] . Thus, the ideal physical process of microscopic image generation simulated by prior methods is not sufficiently capable to model this realistic task.…”

Rapid and label-free detection of gastrointestinal stromal tumor via a combination of two-photon microscopy and imaging analysis

“…1 , we selected the region of interest (ROI) of 1500*1500 pixels for quantitative analysis. For each sample, SHG images as the input images were first filtered by the Frangi filter to enhance the collagen fiber structures from noisy background, and then the enhanced images were segmented into collagen fibers and background by a segmentation algorithm based on Gaussian mixture models [ 13 , 14 ]. The morphological closing and hole-filling were performed to smooth the binary mask of the collagen fibers, and any segment with less than 5 pixels was removed.…”

“…), cross-link space (µm), cross-link density (a.u.) as previously described [ 13 ]. Quantitative results were presented using means with standard deviations (SD).…”