Advances in functional brain imaging now allow sustained rapid 3D visualization of large numbers of neurons inside behaving animals. To decode circuit activity, imaged neurons must be individually segmented and tracked. This is particularly challenging when the brain itself moves and deforms inside a flexible body. The field has lacked general methods for solving this problem effectively. To address this need, we developed a method based on a convolutional neural network (CNN) Awith specific enhancements which we apply to freely moving Caenorhabditis elegans. For a traditional CNN to track neurons across images of a brain with different postures, the CNN must be trained with ground truth (GT) annotations of similar postures. When these postures are diverse, an adequate number of GT annotations can be prohibitively large to generate manually. We introduce ‘targeted augmentation’, a method to automatically synthesize reliable annotations from a few manual annotations. Our method effectively learns the internal deformations of the brain. The learned deformations are used to synthesize annotations for new postures by deforming the manual annotations of similar postures in GT images. The technique is germane to 3D images, which are generally more difficult to analyze than 2D images. The synthetic annotations, which are added to diversify training datasets, drastically reduce manual annotation and proofreading. Our method is effective both when neurons are represented as individual points or as 3D volumes. We provide a GUI that incorporates targeted augmentation in an end-to-end pipeline, from manual GT annotation of a few images to final proofreading of all images. We apply the method to simultaneously measure activity in the second-layer interneurons in C. elegans: RIA, RIB, and RIM, including the RIA neurite. We find that these neurons show rich behaviors, including switching entrainment on and off dynamically when the animal is exposed to periodic odor pulses.