Estimation of in-scanner head pose changes during structural MRI using a convolutional neural network trained on eye...

In-scanner head motion remains a significant obstacle for obtaining high-quality neuroimaging data. This instability introduces systematic errors in cortical thickness and volume measurements derived from structural scans. Few accessible techniques currently exist to monitor these movements during the acquisition process. That uncertainty drove the development of new tracking strategies. Prior research has shown that motion artifacts compromise the validity of brain morphometry. No prior work had resolved the challenge of using existing hardware for this purpose. This gap motivated the exploration of eye-tracking video as a potential source for motion data. The current investigation builds upon these observations to improve data integrity in clinical settings.

The study aims to develop a predictive model that estimates head pose changes during structural MRI. In-scanner motion frequently degrades image quality and biases volumetric brain measurements. Current methods for tracking these movements are often limited or unavailable in standard clinical practice. This uncertainty drove the researchers to leverage existing eye-tracker video for motion monitoring. The investigators sought to create a tool that functions independently of specific imaging parameters. They also intended to avoid the use of external markers attached to the patient. By correlating pose estimates with morphometric data, the team explored the impact of motion on brain imaging results. This work addresses the need for improved motion correction strategies in neuroimaging research.

The research team recruited twenty-one healthy volunteers to participate in the study. Participants performed prompted, stereotyped head movements while undergoing an EPI-BOLD sequence. Simultaneous recordings of eye-tracker video were collected during these sessions. The investigators also obtained T1-weighted whole-brain images for both motion-free and motion-affected conditions. Image coregistration provided the ground truth for head pose changes throughout the scanning duration. This information trained the predictive model to map visual features to specific head orientations. The authors evaluated the model using a hold-out dataset to ensure robust performance. Finally, they correlated these pose estimates with cortical thickness and subcortical volume metrics.

The predictive model demonstrated a significant correlation between video-based estimates and ground truth head pose changes. Increased head motion generally led to a brain-wide reduction in cortical thickness measurements. Some isolated brain regions displayed increased cortical thickness estimates despite the presence of motion. Subcortical volumes were consistently reduced in scans affected by participant movement. These findings confirm the negative impact of motion on standard morphometric outputs. The model successfully operated without the need for physical markers on the patient. This performance was achieved across diverse acquisition parameters, highlighting the flexibility of the approach. The results support the use of this tool for correcting motion-related biases in neuroimaging data.

The researchers propose that their predictive model effectively estimates head pose changes using eye-tracking video. This approach functions independently of specific image acquisition parameters. It requires no physical markers attached to the participant, which enhances its utility in clinical environments. The authors suggest this method improves the neurobiological validity of structural imaging studies. Their findings indicate that motion-affected scans often show reduced cortical thickness and subcortical volumes. The team notes that these pose estimates serve as valuable covariates for future morphometric analyses. This synthesis implies that existing hardware can mitigate common motion-related biases. The study provides a practical tool for researchers working in brain development and disease fields.