To this purpose, most of the available models only employ a pre-selected selection of essential patches or patch clusters as input, perhaps discarding some crucial morphological data, as opposed to using the entire WSIs. In order to properly utilize the whole WSI information, we present a novel survival analysis approach in this work. We demonstrate that using a Vision Transformer (ViT) backbone and the associated convolution operations is a successful strategy for enhancing prediction performance. The most prominent patches and distinctive morphology traits are also identified using a post-hoc manner that is understandable, which improves the model's accuracy and makes it simpler for people to understand the results.
Evaluations on two sizable cancer datasets demonstrate that our suggested model is more useful and easier to understand for predicting survival. Upon approval, we would release the code for general use.

liwen zhang
Institute of Automation, CAS