Segment AnyNeuron

Abstract

Image segmentation plays an integral part in neuroimage analysis and is crucial for understanding brain disorders. Deep Learning (DL) models have shown exponential success in computer vision tasks over the years, including image segmentation. However, to achieve optimal performance, DL models require extensive annotated data for training, which is often the bottleneck to expediting brain-wide image analysis. For segmenting cellular structures such as neurons, the annotation process is cumbersome and time-consuming due to the inherent structural, intensity, and background variations present in the data caused by genetic markers, imaging techniques, etc. We propose an Active Learning-based neuron segmentation framework (Segment AnyNeuron), which incorporates state-of-the-art image segmentation modules - Detectron2 and HQ SAM, and requires minimal ground truth annotation to achieve high precision for brain-wide segmentation of neurons. Our framework can classify and segment completely unseen neuronal data by selecting the most representative samples for manual annotation, thus avoiding the cold-start problem common in Active Learning. We demonstrate the effectiveness of our framework for automated brain-wide segmentation of neurons on a variety of open-source neuron imaging datasets, acquired from different scanners and a variety of transgenic mouse lines.