ATTENTION-BASED DEEP MULTIPLE INSTANCE LEARNING WITH ADAPTIVE INSTANCE SAMPLING

One challenge of training deep neural networks with gigapixel whole-slide images (WSIs) in computational pathology is the lack of annotation at pixel level or regional level due to the high cost and time-consuming labeling effort. Multiple instance learning (MIL) and its attention-based versions are typical weakly supervised learning methods, which allow us to use slide-level labels directly, without the need for pixel or region labels, thus reducing the cost of annotation. However, training a deep neural network with thousands of image regions (patches) per slide is computationally expensive, and it needs a lot of time for convergence. This paper proposes a fast adaptive attention-based deep MIL approach. This approach adaptively selects image regions that are highly predictive of outcome and ignores image regions with little or no information. We empirically show that our proposed approach outperforms the random sampling approach while it is faster than the standard attention-based MIL method (which uses all image regions for training).