DeepBrain: Functional Representation of Neural In-Situ Hybridization Images for Gene Ontology Classification Using Deep Convolutional Autoencoders

Abstract

This paper presents a novel deep learning-based method for learning a functional representation of mammalian neural images. The method uses a deep convolutional denoising autoencoder (CDAE) for generating an invariant, compact representation of in situ hybridization (ISH) images. While most existing methods for bio-imaging analysis were not developed to handle images with highly complex anatomical structures, the results presented in this paper show that functional representation extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Using this CDAE representation, our method outperforms the previous state-of-the-art classification rate, by improving the average AUC from 0.92 to 0.98, i.e., achieving 75% reduction in error. The method operates on input images that were downsampled significantly with respect to the original ones to make it computationally feasible.

Appendix: Network Architecture Description

We provide a brief explanation as to the choice of the main parameters of the CDAE architecture. Our objective was to obtain a more compact feature representation than the 2,004-dimensional vector used in FuncISH. Since a CNN is used, the representation along the grid should capture the two-dimensional structure of the input, i.e., the image dimensions should be determined according to the intended representation vector, while maintaining the aspect ratio of the original input image. Thus, we picked an 1,800-dimensional feature vector, corresponding to an (output) image of size $60 \times 30$. Taking into account the characteristic of max-pooling (i.e., that at each stage the dimension is reduced by 2), the desire to keep the number of layers as small as possible, and the fact that the encoding and decoding phases each contains the same number of layers (resulting in twice the number of layers in the network), we settled for two max-pooling layers, namely an input image of size $240 \times 120$. Between each two max-pooling layers, which eliminate feature redundancy, there is an “array” of 16 convolution layers, each with the purpose of detecting locally connected features from its previous layer. The number of convolution layers (i.e., different filters used) was determined after experimenting with several different layers, all of which gave similar results. Choosing 16 layers (as shown in Fig. 4) provided the best result. We experimented also with various filter sizes for each layer, ranging from $3 \times 3$ to $11 \times 11$; while increasing the filter size significantly increased the amount of network parameters learned, it did not contribute much to the feature extraction or the improvement of the results. Using a learning rate decay in the training of large networks (where there is a large number of randomly generated parameters) has proven helpful in the network’s convergence. Specifically, the combination of a 0.05 learning rate parameter with a 0.9 learning rate decay resulted in an optimal change of the parameter value. In this case, too, small changes in the parameters did not result in significant changes in the results.