AbstractMachine learning approaches are becoming increasingly widespread and are now present in most areas of research. Their recent surge can be explained in part due to our ability to generate and store enormous amounts of data with which to train these models. The requirement for large training sets is also responsible for limiting further potential applications of machine learning, particularly in fields where data tend to be scarce such as developmental biology. However, recent research seems to indicate that machine learning and Big Data can sometimes be decoupled to train models with modest amounts of data. In this work we set out to train a CNN-based classifier to stage zebrafish tail buds at four different stages of development using small information-rich data sets. Our results show that two and three dimensional convolutional neural networks can be trained to stage developing zebrafish tail buds based on both morphological and gene expression confocal microscopy images, achieving in each case up to 100% test accuracy scores. Importantly, we show that high accuracy can be achieved with data set sizes of under 100 images, much smaller than the typical training set size for a convolutional neural net. Furthermore, our classifier shows that it is possible to stage isolated embryonic structures without the need to refer to classic developmental landmarks in the whole embryo, which will be particularly useful to stage 3D culture in vitro systems such as organoids. We hope that this work will provide a proof of principle that will help dispel the myth that large data set sizes are always required to train CNNs, and encourage researchers in fields where data are scarce to also apply ML approaches.Author summaryThe application of machine learning approaches currently hinges on the availability of large data sets to train the models with. However, recent research has shown that large data sets might not always be required. In this work we set out to see whether we could use small confocal microscopy image data sets to train a convolutional neural network (CNN) to stage zebrafish tail buds at four different stages in their development. We found that high test accuracies can be achieved with data set sizes of under 100 images, much smaller than the typical training set size for a CNN. This work also shows that we can robustly stage the embryonic development of isolated structures, without the need to refer back to landmarks in the tail bud. This constitutes an important methodological advance for staging organoids and other 3D culture in vitro systems. This work proves that prohibitively large data sets are not always required to train CNNs, and we hope will encourage others to apply the power of machine learning to their areas of study even if data are scarce.
Precision-Recall versus Accuracy and the Role of Large Data Sets
