The field of medical imaging and interventional navigation is moving towards more accurate and efficient methods for 3D reconstruction, registration, and segmentation. Recent developments have focused on leveraging deep learning techniques, such as vision transformers and neural networks, to improve the precision and robustness of these methods. Additionally, there is a growing interest in developing frameworks that can handle complex anatomical structures, such as blood vessels and spine tissue, and that can provide real-time feedback for interventional procedures. Noteworthy papers include:
- ZeroReg3D, a zero-shot registration pipeline for 3D consecutive histopathology image reconstruction, which achieves accurate 3D reconstruction without requiring large-scale training data.
- Patch2Loc, an unsupervised approach for brain lesion detection that learns to localize patches in structural MRI and detects abnormal patches through location prediction errors.
- Calibrated self-supervised vision transformers for intracranial arterial calcification segmentation, which outperform supervised methods and demonstrate improved robustness to varying slice thicknesses.