One-shot Active Neuron Localization with in vivo Fluorescence Image Sequences

Active neuron localization with in vivo fluorescence image sequences for neuron activity analysis traditionally relies on manual observation. To reduce labor costs, we cast this process as a novel one-shot active neuron localization task requiring minimal user intervention, i.e., only one active neuron template. Observing that neurons switch between active and inactive states over time, we formulate this task as a new one-shot multiple instance learning (MIL) problem, and propose an MIL-based framework integrating multifaceted properties. First, patch proposals are generated heuristically and combined into sequence proposals with temporal profile-based filtering. Second, a deep neural network is trained with self-supervision as the feature extractor. Third, the template feature and the patch feature of sequence proposals are extracted and compared to classify sequences and identify neuron positions in an MIL fashion. Experiments with in vivo data demonstrate our method's clear performance advantage over comparison methods, providing a foundation for future physiological analysis. The code is available at https://github.com/Spritea/OANL.