学术文献库

Multiple instance learning exhibits a powerful approach for whole slide image-based diagnosis in the absence of pixel- or patch-level annotations. In spite of the huge size of hole slide images, the number of individual slides is often rather small, leading to a small number of labeled samples. To improve training, we propose and investigate different data augmentation strategies for multiple instance learning based on the idea of linear interpolations of feature vectors (known as MixUp). Based on state-of-the-art multiple instance learning architectures and two thyroid cancer data sets, an exhaustive study is conducted considering a range of common data augmentation strategies. Whereas a strategy based on to the original MixUp approach showed decreases in accuracy, the use of a novel intra-slide interpolation method led to consistent increases in accuracy.