TL;DR
This paper introduces a novel deep learning AI model that automates the volumetric measurement of vestibular schwannomas (VS) from longitudinal MRI scans after radiosurgery. This approach achieves high accuracy, comparable to expert radiologists, and can track tumor changes over time, regardless of whether the tumor is progressing, regressing, or showing pseudoprogression.
What Problem Does This Paper Address?
Vestibular schwannomas (VS) are tumors that grow on the nerve connecting the ear to the brain. Radiosurgery, like Gamma Knife radiosurgery (GKRS), is a common treatment. However, after GKRS, VS can sometimes swell temporarily (pseudoprogression) before shrinking, making it difficult to assess treatment response accurately. This requires careful, long-term monitoring using MRI scans and precise measurement of tumor volume. Manual tumor segmentation and volume measurement from these scans are time-consuming and can be subjective, leading to potential errors. The paper addresses the need for an automated, accurate, and reliable method to track VS volume changes over time to better guide clinical decisions.
Methods
The core of this work is a proposed "dual-pathway" end-to-end deep learning model, inspired by the U-Net architecture, designed for VS segmentation. This model takes multiple types of MRI scans (T1W+C and T2W) as input. The "dual-pathway" aspect refers to using different convolution kernel sizes (3x3x1 for in-plane and 1x1x3 for through-plane features) in the encoder to better capture information from anisotropic MR images.
The process involves:
- Automated Pre-processing: This includes co-registration of T1W+C and T2W images, slice selection/padding to standardize the number of slices, and cropping to a consistent size.
- Dual-Pathway U-Net Segmentation: The model learns to identify tumor regions by combining features extracted from both T1W+C and T2W images.
- Volume Estimation: The segmented tumor mask is used with voxel size information to calculate the tumor's volume.
The model was trained using a combination of Dice loss and L2 regularization for the initial epochs, and then switched to the Tversky index loss with L2 regularization for the later epochs to better handle false positives and negatives.
Datasets
The study utilized a dataset of 861 patients who underwent Gamma Knife radiosurgery (GKRS) between 1993 and 2008. For training and testing, 1290 MR examinations (T1W+C and T2W parametric MR images) from 381 patients were used. These images were acquired under consistent protocols. The ground truth for training and testing was established by manual delineation of tumor masks by experienced neuroradiologists and neurosurgeons.
Key Insights and Ideas
- Handling Anisotropic MR Images: The dual-pathway approach, with specialized kernels for in-plane and through-plane features, is crucial for accurately segmenting tumors from MR images that have different resolutions along different axes. This is a key innovation for improving segmentation accuracy in this context.
- Robustness to Tumor State: The AI model's ability to accurately measure tumor volume regardless of whether the tumor is progressing, pseudo-progressing, or regressing is a significant advantage. This is particularly important for managing the transient swelling seen in pseudoprogression.
- Combining Modalities: Using both T1W+C and T2W images together allows the model to better delineate tumors, especially those with cystic components or loss of contrast enhancement, which can be challenging for single-modality approaches.
- Clinical Acceptable Accuracy: The AI's measurements are consistently within a range deemed clinically acceptable, often better than the typical 10% error margin for manual measurements.
Key Results
- The proposed dual-pathway model achieved a mean Dice score of 0.90 ± 0.05 on 100 testing data sets, outperforming a single-pathway model (0.87 ± 0.07).
- For longitudinal analysis, the relative volume difference (RVD) between AI-based measurements and expert clinical measurements was minimal across various follow-up time points. The median RVDs ranged from -0.44% to +1.74%, with most points being very close to 0%.
- The AI model's accuracy was reported to be roughly 99% at most follow-up time points.
- The model demonstrated high accuracy in predicting tumor progression or regression trends, as shown in case demonstrations.
Limitations and Future Work
- Study Limitations:
- Retrospective Data: The study relied on retrospective data, which can have inherent limitations.
- Generalizability: The AI model's performance might be higher for VS due to consistent location and contrast intensity compared to other tumor types (e.g., meningioma). Further validation on external datasets and other brain tumors is needed.
- Partial Volume Effects: Amplified upper and lower tumor edges can lead to overestimation of volume.
- Registration Errors: Minor errors in image registration between T1W+C and T2W images, though generally small (around 0.1 mm), can still impact performance.
- Future Work:
- The authors suggest that the proposed AI scheme is applicable to longitudinal follow-up assessments for a variety of therapeutic interventions beyond GKRS for VS.
- They also highlight the growing importance of imaging features for predicting treatment response, suggesting that their accurate volumetric measurements will be increasingly valuable in conjunction with radiomic features.
Summary updated