Peri‐tumoural stroma collagen organization of invasive ductal carcinoma assessed by polarized light microscopy differs between OncotypeDX risk group

Abstract

A commercially available genomic test, OncotypeDX has emerged as a useful postsurgical treatment guide for early stage breast cancer. Despite widespread clinical adoption, there remain logistical issues with its implementation. Collagenous stromal architecture has been shown to hold prognostic value that may complement OncotypeDX. Polarimetric analysis of breast cancer surgical samples allows for the quantification of collagenous stroma abundance and organization. We examine intratumoural collagen abundance and alignment along the tumor-host interface for 45 human samples of invasive ductal carcinoma categorized as low or higher risk by OncotypeDX. Furthermore, we probe the separatory power of collagen alignment patterns to classify unlabeled samples as low or higher OncotypeDX risk group using a linear discriminant (LD) model. No significant difference in mean collagen abundance was found between the two risk groups. However, collagen alignment along the tumor boundary was found to be significantly lower in higher risk samples. The LD model achieved a 71% total accuracy and 81% sensitivity to higher risk samples. Prognostic information extracted from the stromal morphology has potential to complement OncotypeDX as an easyto-implement prescreening methodology.

Machine learning based local recurrence prediction in colorectal cancer using polarized light imaging

“…This study improves upon previous efforts [32][33][34][35][36][37] to create a polarimetric ML workflow adept at providing valuable prognostic insights, by (1) utilizing a wider set of more robust optical polarimetric features and (2) using scalable supervised learning techniques. Specifically, the recent publication of Tumanova et al 37 used Matt-Whitney U-test inferential statistics to demonstrate that certain MM parameters do offer some separation between LR and no-LR groups; however, no predictive analyses were performed.…”

“…This approach allowed the algorithms to focus on the more representative patterns and relationships within the dataset, ultimately enhancing its predictive capabilities (XGBoost results before vs. after elimination of outliers: total accuracy 73% versus 78%, area under the receiver operating characteristic curve (AUROC) 72% versus 77%, sensitivity 31% versus 50%). Consequently, three samples were removed from the study cohort (i.e., leaving 35 . This is all conducted at the ROI-level where the recurrence status is assigned to each region, and it is treated as an individual data point.…”

“…For polarimetric analysis, sample preparation involved chemical dewaxing to avoid possible polarization artifacts. [32][33][34][35][36][37] No further processing was required for polarimetric imaging. Adjacent slides were hematoxylin and eosin (H&E) stained and imaged at 20× magnification on an Aperio ScanScope CS (Leica Biosystems, United States) for the pathologists' region-of-interest (ROI) selection.…”

“…25 More specifically, a PLM technique known as Mueller matrix (MM) polarimetry has been increasingly combined with machine learning (ML) algorithms to directly correlate the underlying biological phenomena with their polarization properties to identify prognostically relevant parameters. [26][27][28][29][30][31] Our recent studies utilized rotating crossed linear polarization microscopy, [32][33][34][35][36] combined with an unsupervised ML clustering pipeline, 32 the latter demonstrating appreciable correlations with 5-year patient survival outcome. 32 However, in that study no associations were investigated with another important clinical endpoint (5-year LR), and the unsupervised learning algorithm used was not easily scalable.…”

Polarimetric biomarkers of peri-tumoral stroma can correlate with 5-year survival in patients with left-sided colorectal cancer

“…Technologically simpler than the Mueller matrix approaches, this methodology has enabled quantitative assessment of stromal maturity and the tumour-stroma ratio in breast cancer tissue samples; comparison against pathologist classifications yielded excellent initial agreement 34 , 35 . Another investigation correlated polarimetrically-derived stromal metrics with genetic prognostication results (OncoTypeDX), also with good agreements 36 . In the current work, we expand significantly on these initial studies by (1) increasing the polarimetric feature space for selecting potential prognostic biomarkers, (2) employing unsupervised clustering algorithms to analyze the data, and (3) correlating the results with actual clinical outcomes (5-year survival).…”