Skip to main content
Fig. 1 | BMC Bioinformatics

Fig. 1

From: InstantDL: an easy-to-use deep learning pipeline for image segmentation and classification

Fig. 1

InstantDL provides an easy to use pipeline for the analysis of biomedical images. a Flow diagram of the pipeline with indicated user action highlighted in gray. (1) One out of four tasks (semantic segmentation, instance segmentation, pixel-wise regression, classification) is selected by the user. Up to twelve parameters can be set in the configuration file to adapt the pipeline to the task. A code snippet illustrates task selection and six of the twelve parameter settings in the configuration file: selected task (“use_algorithm”), path to folder (“path”), if pre-trained weights should be used the path to these (“pretrained_weights_path”) should be set, batch size (“batchsize”) and epochs chosen (“Iterations_Over_Dataset”). (2) Input data is split into train and test sets. The user specifies these by putting the data in the corresponding folders. After executing the python configuration file the pipeline will automatically load the data from the train folder, create a 20 percent validation split, normalize and augment the data (see Methods for details). Training is initiated with either a pre-trained or randomly initialized model. After training, the model predicts test labels: segmentation masks, pixel values or labels for the images in the test set according to the chosen task. (3) Results can be interpreted by the user via statistical and visual assessment of the predicted outcome by comparing it to the ground truth in the test set. b Example output for a 2D semantic segmentation task: Cell nuclei in a brightfield image (left) are segmented with InstantDL (Prediction) using the U-Net, and compared to the original annotation (Groundtruth). The Errormap indicates over- and under-predicted pixels. The image is part of the 2018 Kaggle nuclei segmentation challenge dataset [13]. c Example output for a 2D instance segmentation task (same image as in b): A binary mask is predicted for each object in the image using InstantDLs Mask-RCNN algorithm and compared to the groundtruth. d Example output for a 3D pixel-wise regression task using a U-Net. From stacks of bright-field images (Image) [7] the pipeline predicts a nuclear envelope (Prediction) that resembles the true staining (Groundtruth). The first row shows the x–y-plane, the bottom row the x–z plane of the 3D volume. e Example output for a classification task of benign and leukemic blood cells in blood smears from 200 individuals [17]. We show two exemplary microscopy images (left) of two white blood cell classes, a monoblast and a neutrophil. The white blood cell type is predicted with a ResNet50. The confusion matrix (middle) shows that most of the 15 classes can be well predicted, in accordance to Matek et al. [6]

Back to article page