Scale Embedding Shared Neural Networks for Multiscale Histological Analysis of Prostate Cancer

摘要

In digital pathology, deep learning approaches have been increasingly applied and shown to be effective in analyzingdigitized tissue specimen images. Such approaches have, in general, chosen an arbitrary scale or resolution at which theimages are analyzed for several reasons, including computational cost and complexity. However, the tissuecharacteristics, indicative of cancer, tend to present at differing scales. Herein, we propose a framework that enablesdeep convolutional neural networks to perform multiscale histological analysis of tissue specimen images in an efficientand effective manner. A deep residual neural network is shared across multiple scales, extracting high-level features. Thehigh-level features from multiple scales are aggregated and transformed in a way that the scale information is embeddedin the network. The transformed features are utilized to classify tissue images into cancer and benign. The proposedmethod is compared to other methodologies to combine the feature from different scales. These competing methodscombine the multi-scale features via 1) concatenation 2) addition and 3) convolution. Tissue microarrays (TMAs) wereemployed to evaluate the proposed method and the other competing methods. Three TMAs, including 225 benign and377 cancer tissue samples, were used as training dataset. Two TMAs with 151 benign and 252 cancer tissue samples wasutilized as testing dataset. The proposed method obtained an accuracy of 0.953 and the area under the receiver operatingcharacteristics curve (AUC) of 0.971 (95% CI: 0.955-0.987), outperforming other competing methods. This suggests thatthe proposed multiscale approaches via a shared neural network and scale embedding scheme, could aid in improvingdigital pathology analysis and cancer pathology.