Intraoperative magnetic resonance imaging (iMRI) plays a crucial role in neurosurgery, providing real-time imaging that aids in the maximal safe resection of brain tumours. However, iMRI presents challenges due to its suboptimal imaging conditions, requiring high-speed acquisitions while maintaining diagnostic accuracy. Conventional methods such as compressed sense (CS) reconstruction have been used to enhance image quality, but they often introduce artefacts and reduce the signal-to-noise ratio (SNR).
Deep learning (DL) has emerged as a promising alternative for reconstructing iMRI scans, offering improvements in spatial resolution, image quality and diagnostic confidence. A recent retrospective study published in European Radiology Experimental evaluated the effectiveness of a DL-based reconstruction model compared to conventional CS methods, assessing its potential advantages and limitations in an intraoperative setting.
Methodology and Evaluation of Deep Learning Reconstruction
The study retrospectively analysed imaging data from 40 patients who underwent iMRI during brain tumour resection. The DL model was trained on the fastMRI neuro dataset, designed to simulate the imaging conditions of iMRI while leveraging high-quality training data. Dual surface coils were used for image acquisition and the DL reconstruction was compared to the CS method using multiple quality metrics. Three expert readers—a neurosurgeon and two neuroradiologists—assessed the reconstructed images based on imaging artefacts, perceived spatial resolution, anatomical conspicuity, diagnostic confidence, SNR and contrast. A five-point Likert scale was used to evaluate image quality, and a preference ranking was established to determine the favoured reconstruction technique.
The neuroradiologists assessed improvements in the diagnostic clarity of DL reconstructions, particularly in terms of spatial resolution and contrast, which are vital for identifying residual tumour tissues. The DL reconstructions were evaluated using both qualitative and quantitative assessments, ensuring a comprehensive comparison with CS. Each patient’s imaging data was anonymised and randomly assigned before being reviewed to minimise bias. The readers were given flexibility in choosing their preferred software for evaluation, allowing them to closely examine image details.
Comparison of Image Quality and Expert Preferences
The findings revealed that DL-based reconstruction generally outperformed CS in terms of spatial resolution, anatomic conspicuity and diagnostic confidence. The neuroradiologists strongly preferred DL reconstructions, favouring them in 33 and 39 out of 40 cases. The DL method provided clearer delineation of resection margins and tumour residues, improving overall readability. However, the neurosurgeon preferred the CS reconstruction in 18 out of 40 cases, citing concerns regarding image artefacts and low signal intensity in certain areas. DL-based reconstructions occasionally exhibited striping artefacts and reduced signal strength in regions distant from the receiver coils, potentially impacting intraoperative decision-making. Despite these limitations, DL reconstructions were considered diagnostic in quality and demonstrated potential for further optimisation.
The neuroradiologists noted that the enhanced spatial resolution in DL reconstructions provided a more defined structure of the brain’s anatomy, making it easier to differentiate between tumour margins and surrounding tissues. Additionally, the increased diagnostic confidence afforded by DL was associated with its ability to reduce noise and improve the contrast of intraoperative images. However, the neurosurgeon’s concerns about DL artefacts highlighted the need for refinement to ensure that crucial surgical decisions are not compromised. The disparity in expert opinions underscores the importance of balancing image quality with practical usability in a surgical setting.
Challenges and Future Directions
While DL-based iMRI reconstruction showed promising results, the study highlighted areas for improvement. The low signal intensity in regions far from the receiver coils and occasional striping artefacts were noted as concerns. The variability in expert preferences underscores the importance of tailoring DL-based reconstruction methods to align with the specific needs of both radiologists and neurosurgeons. Future advancements should focus on refining DL algorithms to mitigate artefacts and ensure consistent signal distribution across images. Additionally, expanding the training dataset to include iMRI-specific images could enhance the model’s adaptability to intraoperative imaging conditions. Further research is necessary to validate these findings across larger patient cohorts and multiple clinical sites.
Another consideration is the potential to integrate three-dimensional reconstruction techniques to address the striping artefacts associated with two-dimensional reconstructions. Three-dimensional models could provide improved depth perception and reduce inconsistencies between imaging slices. Additionally, optimising the model to better handle low SNR areas could lead to more reliable intraoperative imaging. Given that iMRI procedures already extend surgical duration, it is crucial that any improvements in image quality do not come at the expense of additional scanning time, as this could increase the risk of infection and surgical complications.
Deep learning-based reconstruction presents a viable alternative to conventional CS methods in intraoperative MRI, demonstrating significant improvements in image quality and diagnostic confidence. Neuroradiologists favoured the DL approach for its enhanced spatial resolution and clarity, while the neurosurgeon’s preference for CS underscores the need for further refinement. Addressing the remaining challenges, such as signal inconsistency and artefacts, will be critical to optimising DL-based iMRI reconstruction for widespread clinical adoption. With continued development, DL has the potential to revolutionise intraoperative imaging, offering real-time, high-quality reconstructions that improve surgical precision and patient outcomes. Further studies are necessary to explore how DL techniques can be optimally integrated into intraoperative workflows while maintaining efficiency and accuracy.
Source: European Radiology Experimental
Image Credit: iStock
