Session Item

Learning survival outcomes on whole slide images (WSI) is a difficult task for deep learning because the size of WSI has both characteristics of extremely large and arbitrary. In this study, we developed spatial pyramid pooling survival networks (SPPSN) to extract risk information from WSIs.

We firstly trained a tissues classifier to segment tissues of gland, stroma, immune cells, and other tissues (adipose, muscle and debris). We also trained a gland formation classifier to segment normal gland tissue and tumor with different gland formation. The two classifiers were designed to extract information that may related to survival outcomes and compress original WSIs into spatial heatmaps with smaller sizes for training. Next, we used spatial pyramid pooling (SPP) to extract SPP features from spatial heatmaps and then applied a deep survival networks after SPP features. The total architecture was called SPPSN and was demonstrated in Fig1. SPPSN can obtain deep survival grade form WSIs and the grade was evaluated by c index and Kaplan-Meier curves.

Fig1. (A) is the work flow of developing gland formation and tissue category classifiers. (B) is the work flow of calculating differentiation grade and deep survival grade. 

The training process was performed on TCGA COAD dataset and TCGA READ dataset. The combination of COAD and READ dataset was randomly split into training set (923 WSIs) and validation set (230 WSIs). A local institutional dataset with 108 WSIs was enrolled as external test set. In validation set and test set, c index of deep survival grade from SPPSN were 0.64 and 0.64, respectively. In Kaplan-Meier curves, WSIs were split into low risk group and high risk group by median cut off point. The log rank test of Kaplan-Meier curves (Fig2) was performed on validation set and test set, the p values were both lower than 0.05.

Fig2 Kaplan-Meier curves of SGFR and deep survival grade in validation set and test set.

The SPPSN can successfully extract risk information from WSIs for survival outcomes prediction. Our method provide a new way to handle WSIs and create a new predictor which may potentially benefit therapy decision making for colorectal cancer.