BACKGROUND: Histopathology image analysis is a gold standard for cancer recognition and diagnosis. Automatic analysis of histopathology images can help pathologists diagnose tumor and cancer subtypes, alleviating the workload of pathologists. There are two basic types of tasks in digital histopathology image analysis: image classification and image segmentation. Typical problems with histopathology images that hamper automatic analysis include complex clinical representations, limited quantities of training images in a dataset, and the extremely large size of singular images (usually up to gigapixels). The property of extremely large size for a single image also makes a histopathology image dataset be considered large-scale, even if the number of images in the dataset is limited. RESULTS: In this paper, we propose leveraging deep convolutional neural network (CNN) activation features to perform classification, segmentation and visualization in large-scale tissue histopathology images. Our framework transfers features extracted from CNNs trained by a large natural image database, ImageNet, to histopathology images. We also explore the characteristics of CNN features by visualizing the response of individual neuron components in the last hidden layer. Some of these characteristics reveal biological insights that have been verified by pathologists. According to our experiments, the framework proposed has shown state-of-the-art performance on a brain tumor dataset from the MICCAI 2014 Brain Tumor Digital Pathology Challenge and a colon cancer histopathology image dataset. CONCLUSIONS: The framework proposed is a simple, efficient and effective system for histopathology image automatic analysis. We successfully transfer ImageNet knowledge as deep convolutional activation features to the classification and segmentation of histopathology images with little training data. CNN features are significantly more powerful than expert-designed features.

Recently, [Li, Nguyen, Woodruff, STOC 2014] showed any 1-pass constant probability streaming algorithm for computing a relation f on a vector x in {-m, -(m-1), ..., m}^n presented in the turnstile data stream model can be implemented by maintaining a linear sketch Ax mod q, where A is an r times n integer matrix and q = (q_1, ..., q_r) is a vector of positive integers. The space complexity of maintaining Ax mod q, not including the random bits used for sampling A and q, matches the space of the optimal algorithm. We give multiple strengthenings of this reduction, together with new applications. In particular, we show how to remove the following shortcomings of their reduction: 1. The Box Constraint. Their reduction applies only to algorithms that must be correct even if x_{infinity} = max_{i in [n]} |x_i| is allowed to be much larger than m at intermediate points in the stream, provided that x is in {-m, -(m-1), ..., m}^n at the end of the stream. We give a condition under which the optimal algorithm is a linear sketch even if it works only when promised that x is in {-m, -(m-1), ..., m}^n at all points in the stream. Using this, we show the first super-constant Omega(log m) bits lower bound for the problem of maintaining a counter up to an additive epsilon*m error in a turnstile stream, where epsilon is any constant in (0, 1/2). Previous lower bounds are based on communication complexity and are only for relative error approximation; interestingly, we do not know how to prove our result using communication complexity. More generally, we show the first super-constant Omega(log(m)) lower bound for additive approximation of l_p-norms; this bound is tight for p in [1, 2]. 2. Negative Coordinates. Their reduction allows x_i to be negative while processing the stream. We show an equivalence between 1-pass algorithms and linear sketches Ax mod q in dynamic graph streams, or more generally, the strict turnstile model, in which for all i in [n], x_i is nonnegative at all points in the stream. Combined with [Assadi, Khanna, Li, Yaroslavtsev, SODA 2016], this resolves the 1-pass space complexity of approximating the maximum matching in a dynamic graph stream, answering a question in that work. 3. 1-Pass Restriction. Their reduction only applies to 1-pass data stream algorithms in the turnstile model, while there exist algorithms for heavy hitters and for low rank approximation which provably do better with multiple passes. We extend the reduction to algorithms which make any number of passes, showing the optimal algorithm is to choose a new linear sketch at the beginning of each pass, based on the output of previous passes.

We propose a simple, efficient and effective method using deep convolutional activation features (CNNs) to achieve stat- of-the-art classification and segmentation for the MICCAI 2014 Brain Tumor Digital Pathology Challenge. Common traits of such medical image challenges are characterized by large image dimensions (up to the gigabyte size of an image), a limited amount of training data, and significant clinical feature representations. To tackle these challenges, we transfer the features extracted from CNNs trained with a very large general image database to the medical image challenge. In this paper, we used CNN activations trained by ImageNet to extract features (4096 neurons, 13.3% active). In addition, feature selection, feature pooling, and data augmentation are used in our work. Our system obtained 97.5% accuracy on classification and 84% accuracy on segmentation, demonstrating a significant performance gain over other participating teams.