9+！ Predicting lymph node status from histology of colorectal cancer by deep learning

Introduction

Lymph node status is a prognostic marker , For colorectal cancer (CRC) Treatment decisions have a big impact .

Background introduction

In recent years, deep learning has been a research hotspot , This article is brought to you by Xiaobian today , Whether the image features extracted from routine histological sections and clinical data by deep learning model can be used to predict CRC Lymph node metastasis (LNM). Published in 《European Journal of Cancer》 On , The influence factor is 9.162, The title of the article is Deep learning can predict lymph node status directly from histology in colorectal cancer.

Data is introduced

The patient cohort used in model training and internal validation in this study was from DACHS （Darmkrebs: Chancen der Verhu¨tung durch Screening） Research , The data set used for external validation is from TCGA. This includes those from DACHS Studied 2003 - 2014 Diagnosed in 2431 Patients and from TCGA Studied 1998 - 2013 Annual diagnosis 582 Famous patients （ chart 1）.

chart 1

Result analysis

01

Slide based AI predictor （SBAIP） Construction and performance of

The in-depth learning framework of this study is selected in Nvidia Clara Of CAMELYON16 In the challenge H&E Pre trained convolutional neural networks on colored slides (CNN) ResNet18, among , stay CNN A linear classifier is trained on the extracted tile features , And average the scores of all tiles , To get a single score for each patient . A simplified overview of the pipeline for this study is shown in Figure 2 Shown .

chart 2

As shown in Figure 3 , The image classifier has achieved 71.0% Of AUROC. On an external test set ,AUROC by 61.2%.

chart 3

02

Construction and performance of clinical classifier

As “ Clinical baseline ”, This study is based on the patient's age 、 Gender 、T by stages 、 Tumor location （ The colon / Rectum ） And laterality （ Distal / Proximal end ） The data were analyzed by logistic regression , To achieve maximum data availability and accuracy for the first study , This study ultimately used the pathology of colon and rectal cancer T Staging information （pT by stages ） Not clinical T by stages （cT by stages ）.

As shown in Figure 4 , This study is based solely on the following T Clinical classifiers for patient data including staging have produced... On the internal test set 67.0% Of AUROC, Similar or even better results have been achieved on the external test set AUROC 71.1%, It shows that the performance of clinical classifier is very robust .

chart 4

03

Combined classifiers （SBAIP+ Clinical classifier ） Construction and performance of

In this study, a logistic regression model is established , Contains the same patient data as in the previously described clinical classifier , And additionally integrated SBAIP Patient score , The clinical classifier and clinical data are compared with SBAIP The combination of scores is established L1 Regularization logistic The regression model (“clinical”、“clinical +SBAIP”、“ClinicalwoT”、“ClinicalwoT+SBAIP”、“T stage+SBAIP”)..

When using combined classifiers , This study achieved the best performance . On the internal test set , Combined model AUROC by 74.1%（ surface 1, chart 3）, Its external performance is equivalent to that of clinical classifier （AUROC 70.5%）, Omit... From patient data T The stage information will cause the classifier (clinicalwoT) Significant performance degradation on both test sets .

surface 1

Editor's summary

This study used the recent popular deep learning to build a lymph node prediction model for colon cancer , Although the result is AUC The value is not high , But because the idea is novel , Still achieved 9+ Height . The data and ideas used in this study are worthy of our reference in future research ！