Learning Cell Nuclei Segmentation Using Labels Generated With Classical Image Analysis Methods

Creating manual annotations in a large number of images is a tedious bottleneck that limits deep learning use in many applications. Here, we present a study in which we used the output of a classical image analysis pipelineas labels when training a convolutional neural network(CNN). This may not only reduce the time experts spend annotating images but it may also lead to an improvement of results when compared to the output from the classical pipeline used in training. Inour application, i.e.,cell nuclei segmentation,we generated the annotations using CellProfiler(a tool for developing classical image analysis pipelines for biomedical applications)and trained on them a U-Net-based CNN model. The best model achieved a 0.96 dice-coefficient of the segmented Nuclei and a 0.84 object-wise Jaccard indexwhich was better than the classical method used for generating the annotations by 0.02and 0.34, respectively. Our experimental results show that in this application, not only such training is feasiblebut also thatthe deep learning segmentationsare a clear improvement compared to the output from the classical pipelineused for generating the annotations.