Multi-scale Intelligent Quantification And Prognosis Analysis Of Tumor Microenvironment In Colorectal Cancer Based On Digital Pathological Images

As the third most common cancer in the world and China,colorectal cancer(CRC)has a high morbidity and mortality rate,posing a serious threat to people's health.Tumor microenvironment plays a key role in the diagnosis,treatment and prognosis of CRC.It is a difficult and hot issue to accurately quantify tumor microenvironment(TME)information in the current field of cancer research.As the gold standard of tumor diagnosis,histopathology slides provide materials for accurate quantification of TME.Whole-slide images(WSIs)can provide TME information at different scales from tissue to cell,such as tumor-stroma ratio,tumor budding,and tumor infiltrating lymphocytes.In recent years,the emerging digital image processing technology,including artificial intelligence,provides a reliable and powerful tool for the intelligent quantification of TME information in whole-slide images.This paper aimed to quantify TME of colorectal cancer with objective and intelligent pathology image analysis.Different digital image processing methods,including classification and segmentation,were used to quantify TME information at tissue and cell levels.Several quantified TME-related factors,such as tumor-stroma ratio,Crohn's-like lymphoid reaction(CLR)density,and Immunoscore,were associated with the survival outcome of CRC patient.With the advantages of fast computation,high repeatability and automatic,digital pathological image analysis has great potential for mining and quantifying the TME information of CRC,thus promoting the clinical application of quantitative TME-related indicators.The main research contents and innovations of this paper are as follows:1.For eliminating the manual annotation effort and speeding up calculation process in the WSI analysis,this paper proposed a fully automatic region of interest(ROI)identification pipeline using low-magnification immunohistochemical WSI.The constructed Hist-Immune signature was a TNM stage-independent prognostic factor,and the complete model containing the signature had better prognostic performance than the clinical model(C index,0.727 vs.0.694).By combining the color deconvolution and global threshold segmentation,the pipeline could identify the immune status of CRC patients without ROI annotation from slide level.2.For reducing the sample size of the labeled dataset in the deep learning training process,an active learning algorithm was proposed,which used a first-easy-then-hard strategy.In other words,the deep learning model was continuously fine-tuned through the maximum confidence least confidence strategy,which greatly reduced the sample size required for convolutional neural network training and quantified 9 CRC tissue types(compared with the random strategy,active learning reduced the labeling amount by about 90%).Next,several TME-related prognostic biomarkers such as tumor-stroma ratio,mucous tumor ratio,and CLR density,were calculated from WSIs from tissue level.3.For reducing the influence of immunohistochemical stain degree on image analysis,a step-by-step local threshold segmentation algorithm was proposed.An Immunoscore system was built by quantifying the density of CD3~+and CD8~+T cells infiltration within the core of tumor and tumor invasion margin on WSIs.The complete model with Immunoscore included had better prognostic performance than the referrence model(i AUC,0.751 vs.0.708).The Immunoscore could classify patients with colon cancer into different risk stratification from cellular level.4.For combining the TME information from different dimensions,a multi-scale automated analysis pipeline for WSIs was proposed.A prognostic index,the stroma-immune score,was calculated by quantifying immune cells infiltration in the stroma region.The deep learning model was re-trained by transfer learning,and the accuracy of 97.3%was achieved in the test set.The constructed stroma-immune score combined with TNM staging had better prognostic performance than TNM staging alone(C index,0.70 vs.0.62,P<0.001).The multi-scale TME information was quantifying from tissue level to cell level.In short,this paper aimed to automatically quantify multi-scale TME information from WSIs.Digital pathology image analysis methods,such as the traditional segmentation algorithm and deep learning,were used to quantify TME-related prognostic biomarkers from WSIs at different levels.There biomarkers revealed the important role of TME in the diagnosis,treatment and prognosis of CRC patients.