Unsupervised Imbalanced Registration for Enhancing Accuracy and Stability in Medical Image Registration.

BACKGROUND: Medical image registration plays an important role in several applications. Existing approaches using unsupervised learning encounter issues due to the data imbalance problem, as their target is usually a continuous variable. OBJECTIVE: In this study, we introduce a novel approach known as Unsupervised Imbalanced Registration, to address the challenge of data imbalance and prevent overconfidence while increasing the accuracy and stability of 4D image registration. METHODS: Our approach involves performing unsupervised image mixtures to smooth the input space, followed by unsupervised image registration to learn the continual target. We evaluated our method on 4D-Lung using two widely used unsupervised methods, namely VoxelMorph and ViT-V-Net. RESULTS: Our findings demonstrate that our proposed method significantly enhances the mean accuracy of registration by 3%-10% on a small dataset while also reducing the accuracy variance by 10%. CONCLUSION: Unsupervised Imbalanced Registration is a promising approach that is compatible with current unsupervised image registration methods applied to 4D images.

Dlung: Unsupervised Few-Shot Diffeomorphic Respiratory Motion Modeling.

Lung image registration plays an important role in lung analysis applications, such as respiratory motion modeling. Unsupervised learning-based image registration methods that can compute the deformation without the requirement of supervision attract much attention. However, it is noteworthy that they have two drawbacks: they do not handle the problem of limited data and do not guarantee diffeomorphic (topology-preserving) properties, especially when large deformation exists in lung scans. In this paper, we present an unsupervised few-shot learning-based diffeomorphic lung image registration, namely Dlung. We employ fine-tuning techniques to solve the problem of limited data and apply the scaling and squaring method to accomplish the diffeomorphic registration. Furthermore, atlas-based registration on spatio-temporal (4D) images is performed and thoroughly compared with baseline methods. Dlung achieves the highest accuracy with diffeomorphic properties. It constructs accurate and fast respiratory motion models with limited data. This research extends our knowledge of respiratory motion modeling.

Transformer-Based High-Frequency Oscillation Signal Detection on Magnetoencephalography From Epileptic Patients.

High-frequency oscillations (HFOs), observed within 80-500 Hz of magnetoencephalography (MEG) data, are putative biomarkers to localize epileptogenic zones that are critical for the success of surgical epilepsy treatment. It is crucial to accurately detect HFOs for improving the surgical outcome of patients with epilepsy. However, in clinical practices, detecting HFOs in MEG signals mainly depends on visual inspection by clinicians, which is very time-consuming, labor-intensive, subjective, and error-prone. To accurately and automatically detect HFOs, machine learning approaches have been developed and have demonstrated the promising results of automated HFO detection. More recently, the transformer-based model has attracted wide attention and achieved state-of-the-art performance on many machine learning tasks. In this paper, we are investigating the suitability of transformer-based models on the detection of HFOs. Specifically, we propose a transformer-based HFO detection framework for biomedical MEG one-dimensional signal data. For signal classification, we develop a transformer-based HFO (TransHFO) classification model. Then, we investigate the relationship between depth of deep learning models and classification performance. The experimental results show that the proposed framework outperforms the state-of-the-art HFO classifiers, increasing classification accuracy by 7%. Furthermore, we find that shallow TransHFO ( < 10 layers) outperforms deep TransHFO models (≥10 layers) on most data augmented factors.