Survival Prediction via Deep Attention-Based Multiple-Instance Learning Networks with Instance Sampling

Survival prediction via training deep neural networks with giga-pixel whole-slide images (WSIs) is challenging due to the lack of time annotation at the pixel level or patch (instance). Multiple instance learning (MIL), as a typical weakly supervised learning method, aims to resolve this challenge by using only the slide-level time. The attention-based MIL method leverages and enhances performance by weighting the instances based on their contribution to predicting the outcome. A WSI typically contains hundreds of thousands of image patches. Training a deep neural network with thousands of image patches per slide is computationally expensive and time-consuming. To tackle this issue, we propose an adaptive-learning strategy where we sample a subset of informative instances/patches more often to train the deep survival neural networks. We also present other sampling strategies and compare them with our proposed sampling strategy. Using both real-world and synthesized WSIs for survival, we show that sampling strategies significantly can significantly reduce computing time while result in no or negligible performance loss. We also discuss the benefits of each instance sampling strategy in different scenarios