Efficient Multiple Instance Convolutional Neural Networks for Gigapixel Resolution Image Classification

Convolutional Neural Networks (CNNs) are state-of-theart models for many image and video classification tasks. However, training on large-size training samples is currently computationally impossible. Hence when the training data is multi-gigapixel images, only small patches of the original images can be used as training input. Since there is no guarantee that each patch is discriminative, we advocate the use of Multiple Instance Learning (MIL) to combine evidence from multiple patches sampled from the same image. In this paper we propose a framework that integrates MIL with CNNs. In our algorithm, patches of the images or videos are treated as instances, where only the imageor video-level label is given. Our algorithm iteratively identifies discriminative patches in a high resolution image and trains a CNN on them. In the test phase, instead of using voting to the predict the label of the image, we train a logistic regression model to aggregate the patch-level predictions. Our method selects discriminative patches more robustly through the use of Gaussian smoothing. We apply our method to glioma (the most common brain cancer) subtype classification based on multi-gigapixel whole slide images (WSI) from The Cancer Genome Atlas (TCGA) dataset. We can classify Glioblastoma (GBM) and Low-Grade Glioma (LGG) with an accuracy of 97%. Furthermore, for the first time, we attempt to classify the three most common subtypes of LGG, a much more challenging task. We achieved an accuracy of 57.1% which is similar to the inter-observer agreement between experienced pathologists.