Automatic Annotation Of Drosophila Embryonic Images Based On Convolutional Neural Networks

The Drosophila embryonic gene expression images provide important spatiotemporal expression information for understanding the mechanisms of Drosophila embryogenesis.They are helpful for revealing the relavant gene functions,interactions and regulatory networks during Drosophila embryogenesis.With the rapid increasement of Drosophila embryonic gene expression images,it is imperative to develop automatic annotation tools since manual annotation is very labor-intensive and also requires professional knowledge.However,automatic annotation of these images is an challenging task because of the following reason.Unlike the auto-annotation for natural images,the corresponding labels(terms from a controlled vocabulary)of gene expression images are assigned to genes rather than images.Each gene corresponds to a set of images and different genes may be associated with different number of labels.In addition,the annotation terms spatially correspond to local expression patterns of images,yet they are assigned collectively to groups of images and the relations between the terms and the image regions are unknown.Thus,conventional machine learning methods are not applicable in this scenario.In this study,we treat this task as a multi-instance multi-label learning problem,and propose a hierachical multi-instance multi-label learning framework,called HMIML.We implement HMIML at image-level and gene-level respectively,both using convolutional neural networks(CNN).In addition,an image stitching strategy is employed to get a combined image feature representation at gene-level.To be specific,we use a pre-trained VGG16 model to obtain feature representation,and we design a new deep model to train these representations.Moreover,we adopt some effective strategies to enhance the model performance.Experimental results on the FlyExpress database show that HMIML enhances annotation accuracy on all developmental stages compared with the existing methods.