Online

ESTRO 2020

Session Item

Clinical track: Lower GI (colon, rectum, anus)
9306
Poster
Clinical
23:00 - 23:00
Predicting the consensus molecular subtypes of colorectal cancer by weakly supervised deep learning
PO-1085

Abstract

Predicting the consensus molecular subtypes of colorectal cancer by weakly supervised deep learning
Authors: CHEN|, Shenlun(1)*[schrodinger999@gmail.com];Zhen|, Zhang(1);Wang|, Jiazhou(1);Hu|, Weigang(1);
(1)Fudan University Shanghai Cancer Center, Department of Radiation Oncology, Shanghai, China;
Show Affiliations
Purpose or Objective

The consensus molecular subtypes (CMS) of colorectal cancer is the valuable reference for therapy. However, CMS is hardly applied in clinical routine since the requirement of gene sequencing which is time-consuming and expensive. In this study we manage to predict CMS using weakly supervised deep learning on whole slide images (WSI).

Material and Methods

The WSIs enrolled into this study contains The Cancer Genome Atlas (TCGA) COAD project and TCGA-READ project. The former project was used for training and testing while the latter project was used for validation. The training set contains 780 WSIs, the testing set contains 6 WSIs and validation set contain 302 WSIs. First, we randomly cropped 1024 patches with shape of 299 * 299 pixels from each training WSI and predicted these slices through a deep learning classification model with initial random weight. For each WSI, 4 patches with highest probability of matched CMS were selected for model training. The weight of deep learning classification model will be update after one training circle. The whole training process was run 50 times. Heatmap of each WSIs were generated by sliding window technique in validation set. A single heatmap would be spited into 25 grids and CMS score will decide on each grid. The CMS for WSI will decide base on the results of all grid. Fig.1 gives an example of CMS decision.

Results

The accuracy of CMS prediction in validation set archived to 73%. Heatmaps were generated from WSIs in testing sets and Fig.2 demonstrate an example. The heatmaps fit the clinical illustration of CMS, for example, CMS2 and CMS3 are highly related to epithelial tissues and CMS4 is highly related to stromal tissues.

Conclusion

Weak supervised deep learning method have potential to predict CMS on WSIs. This model can predict CMS on WSIs directly and avoid gene sequencing which may make it possible for CMS becoming an accessible reference for colorectal cancer’s therapy.