Joint Appearance-Feature Domain Adaptation: Application to QSM Segmentation Transfer

Abstract

Quantitative susceptibility mapping (QSM) is a magnetic resonance imaging technique used to quantitatively measure the iron content in the brain. Patients with Parkinson’s disease are reported having increased iron deposition, especially in substantia nigra (SN) which is a relatively small gray matter structure located in the midbrain. The automatic segmentation of SN is a critical prerequisite step to facilitate the progression of evaluating the course of Parkinson’s disease. However, the imaging protocol and reconstruction methods in QSM acquisition vary largely, rendering great challenges in constructing and applying image segmentation models. Thus, a model trained on a certain dataset often performs poorly on datasets from other scanners or reconstruction methods. To quickly transfer a trained segmentation model to a dataset acquired in a new instrument, we have developed a joint appearance-feature domain adaptation framework (JAFDAF) to transfer the knowledge from the source to the target domains for improved SN segmentation. In particular, we perform domain adaption in both appearance and feature spaces. In the appearance space, we use region-based histogram matching and a neural network to align the grayscale ranges of images between these two domains. In the feature space, we propose a domain regularization layer (DRL) by utilizing the idea of neural architecture search (NAS) to enforce the convolution kernels for learning features that are efficacious in both domains. Ablation experiments have been carried out to evaluate the proposed JAFDAF framework, and the experimental results on 27 subjects show that our method achieves up to 12% over the baseline model and about 5% over a fine-tuning approach.