Quantitative Histopathology of Stained Tissues using Color Spatial Light Interference Microscopy (cSLIM)

Kevin W Eliceiri,Hassaan Majeed,Andre Kajdacsy-Balla,Adib Keikhosravi,Mikhail E Kandel,Yuming Liu,Tan H Nguyen,Krishnarao Tangella,Gabriel Popescu

doi:10.1038/s41598-019-50143-x

Abstract

Tissue biopsy evaluation in the clinic is in need of quantitative disease markers for diagnosis and, most importantly, prognosis. Among the new technologies, quantitative phase imaging (QPI) has demonstrated promise for histopathology because it reveals intrinsic tissue nanoarchitecture through the refractive index. However, a vast majority of past QPI investigations have relied on imaging unstained tissues, which disrupts the established specimen processing. Here we present color spatial light interference microscopy (cSLIM) as a new whole-slide imaging modality that performs interferometric imaging on stained tissue, with a color detector array. As a result, cSLIM yields in a single scan both the intrinsic tissue phase map and the standard color bright-field image, familiar to the pathologist. Our results on 196 breast cancer patients indicate that cSLIM can provide stain-independent prognostic information from the alignment of collagen fibers in the tumor microenvironment. The effects of staining on the tissue phase maps were corrected by a mathematical normalization. These characteristics are likely to reduce barriers to clinical translation for the new cSLIM technology.

Highlights

Likelihood scores for each EC, generated by the classifier from the three validation trials, were pooled together[70] to generate an overall ROC curve [Fig. 4(i)] and the AUC was used as a metric for classifier accuracy
The results were extracted using an open source MATLAB-based tool called CurveAlign, algorithmic details of which have already been described in a number of publications[13,54,71,72]
ECs were segmented out from all the core images so that the cellular structures within them did not interfere with the process of fiber extraction during curvelet transformation

Summary

Introduction

Using pathologist diagnosis result for each EC as the class label (benign or malignant) and its overall feature vector as the predictor, an LDA classifier was trained. Feature vectors for unknown ECs were input to the classifier which generated likelihood scores for the output classes. Likelihood scores for each EC, generated by the classifier from the three validation trials, were pooled together[70] to generate an overall ROC curve [Fig. 4(i)] and the AUC was used as a metric for classifier accuracy.

Results

Conclusion