Automatic MRI Lymph Node Annotation From CT Labels

After annotating a medical imaging modality that is relatively straightforward to label, doctors often expect automatic annotations for images from other modalities of the same region, even though these modalities differ in contrast and structure. This study focuses on creating automatic lymph node annotation in MRI images using available CT annotations via deep-learning models. Training such models typically requires partial MRI labels for semi-supervision. However, annotating lymph nodes in MRI images is particularly challenging due to their small size and the high cost of MRI scans. These factors make it difficult to create labeled MRI datasets for deep learning model training. Moreover, existing cross-modal annotation methods primarily focus on large tumors and require large datasets, making them unsuitable for small lymph nodes with less training data. We address these challenges using cross-modal supervision through image registration. Our algorithm reduces the burden of manual annotation and the reliance on large labeled datasets and eliminates the need for any MRI ground truth. The algorithm has three steps: 1) unsupervised deformable image translation-based registration of MRI to CT image, producing registered MRI; 2) annotating lymph nodes in registered MRI with the available CT labels; and 3) deregistration of registered annotated MRI back to the original shape of MRI. The translation-based registration model for the algorithm’s first and third steps uses a discriminator-free StyleGAN2 translation network and a deformable image registration network with a U-Net-inspired architecture. This registration network includes local and global feature extraction modules, a local-global spatial correlation module, and a superresolution loss function. Our approach eliminates the need for MRI labels by registering MRI with CT images. Experiments show 2.19% and 4.08% MSE reductions, 5.40% and 3.28% SSIM improvements, 29.85% and 3.82% NCC increases for cross-modality and mono-modality registration, respectively, along with a 36.7% training speedup over state-of-the-art translation-based registration models. The lymph node annotation method achieves an average of 74.3% DSC in the region of interest. It also has broader applications in multimodality image segmentation. We open-source the code through a GitHub repository.INDEX TERMS Image registration, image annotation, magnetic resonance imaging, computed tomography, unsupervised learning, superresolution.