Content-consistent adaptation network for substantia nigra segmentation in cross-modalities and -datasets scenarios without retraining

Fully automated segmentation of substantia nigra (SN) is crucial for developing the interpretable computer-aided diagnosis system for Parkinson’s disease (PD). In real-world applications, the segmentation models may be trained with high quality medical images, while the testing dataset may be scanned with an older machine and is of relatively lower quality. To address this domain shift problem, conventional domain adaptation methods propose various strategies to adapt the segmentation model to the testing data. These conventional adaptation methods typically include a retraining stage in the target domain, either for the segmentation model or a translation model. This process makes the predictions in real multi-modalities and -datasets scenarios highly time-consuming. To enable rapid SN segmentation in cross-modalities and -datasets tasks, this paper proposes a content-consistent adaptation network using only source domain data to synthesize training data with the consistent contents of the target domain. The proposed method requires no retraining during the testing stage, thus enabling rapid SN segmentation across modalities and datasets. The results indicated the competitive performances of our proposed method in both the cross-modalities and -datasets tasks.