Multi-Source Unsupervised Domain Adaptation for Neuron Membrane Segmentation via Feature Enhancement

Neuronal membrane segmentation is an important step in neuroscience research, and its automation can significantly improve research efficiency. However, neuronal boundary segmentation in electron microscopy (EM) images always suffers from the lack of sufficient manual labels. Unsupervised domain adaptation (UDA) methods are used to improve the generalization capabilities of typical deep learning frameworks for medical image analysis by leveraging unlabeled data. Compared with the single-source unsupervised domain adaptation (SUDA) methods, the multi-source unsupervised domain adaptation (MUDA) methods can further improve the performance of the framework in the target domain by leveraging knowledge from multiple source domains. In this paper, we propose an MUDA-based approach for neuronal membrane segmentation in EM images. In detail, our approach first introduces a multi-source feature fusion strategy to converge information from different source domains, improving performance on the target domain. Then, to provide more effective features from a single source domain to assist multi-source domain adaptation, we design multiple intermediate domains for the segmentation task. Finally, we implement a multi-stage strategy to make full use of the intermediate domains. Experiments on three public neuronal membrane datasets validated the superiority of our proposed method.