Abstract
BackgroundDeep learning techniques have been successfully applied to bioimaging problems; however, these methods are highly data demanding. An approach to deal with the lack of data and avoid overfitting is the application of data augmentation, a technique that generates new training samples from the original dataset by applying different kinds of transformations. Several tools exist to apply data augmentation in the context of image classification, but it does not exist a similar tool for the problems of localization, detection, semantic segmentation or instance segmentation that works not only with 2 dimensional images but also with multi-dimensional images (such as stacks or videos).ResultsIn this paper, we present a generic strategy that can be applied to automatically augment a dataset of images, or multi-dimensional images, devoted to classification, localization, detection, semantic segmentation or instance segmentation. The augmentation method presented in this paper has been implemented in the open-source package CLoDSA. To prove the benefits of using CLoDSA, we have employed this library to improve the accuracy of models for Malaria parasite classification, stomata detection, and automatic segmentation of neural structures.ConclusionsCLoDSA is the first, at least up to the best of our knowledge, image augmentation library for object classification, localization, detection, semantic segmentation, and instance segmentation that works not only with 2 dimensional images but also with multi-dimensional images.
Highlights
Deep learning techniques have been successfully applied to bioimaging problems; these methods are highly data demanding
We present a generic method, see “Methods” section, that can be applied to automatically augment a dataset of images devoted to classification, localization, detection, semantic segmentation, and instance segmentation using the classical image augmentation transformations applied in object recognition; this method can be applied to multi-dimensional images
In the case of object classification, each image is labeled with a prefixed category; for object localization, the position of the object in the image is provided using the bounding box; for object detection, a list of bounding boxes and the category of the objects inside those boxes are given; in semantic segmentation, each pixel of the image is labeled with the class of its enclosing object; and, in instance segmentation, each pixel of the image is labeled with the class of its enclosing object and objects of the same class are distinguished among them
Summary
Deep learning techniques have been successfully applied to bioimaging problems; these methods are highly data demanding. A successful method that has been applied to deal with the problem of limited amount of data is data augmentation [7, 8] This technique consists in generating new training samples from the original dataset by applying transformations that do not alter the class of the data. This method has been successfully applied in several contexts such as brain electron microscopy image segmentation [9], melanoma detection [3], or the detection of gastrointestinal diseases from endoscopical images [5]. Several libraries, like Augmentor [10] or Imgaug [11], Casado-García et al BMC Bioinformatics (2019) 20:323 and deep learning frameworks, like Keras [12] or Tensorflow [13], provide features for data augmentation in the context of object classification
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
