Interinstitutional Portability of a Deep Learning Brain MRI Lesion Segmentation Algorithm.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Neuroscience
• Radiology and Diagnostic Imaging

Background:

• Brain MRI lesion segmentation algorithms are crucial for diagnosis and treatment planning.
• Performance of algorithms trained at one institution often degrades when applied to data from another due to variations in imaging protocols and patient populations.
• Multi-institutional training datasets are explored to enhance algorithm generalizability.

Purpose of the Study:

• To evaluate the performance of a brain MRI lesion segmentation algorithm trained at one institution (IN1) when tested at a second institution (IN2).
• To assess the impact of incorporating multi-institutional training data on mitigating performance decline in external validation.
• To determine the optimal strategy for augmenting training datasets to improve cross-institutional generalizability.

Main Methods:

• A 3D U-Net model was initially trained on 293 brain MRI scans from IN1 and tested on 51 scans from IN2.
• The model was subsequently retrained with additional datasets: 285 multi-institution tumor segmentations, 198 IN2 tumor segmentations, and 34 IN2 heterogeneous lesion segmentations.
• Segmentation performance was quantified using Dice coefficients on both IN1 and IN2 test datasets.

Main Results:

• The algorithm demonstrated accurate segmentation across various brain pathologies but showed reduced performance on external data (median Dice score 0.70 for IN2 vs. 0.76 for IN1).
• Adding 483 cases of a single pathology, even from IN2, did not significantly improve external performance (median Dice 0.72, P = .10).
• Incorporating a small subset (34 cases) of diverse IN2 lesion data (10% of total training data) significantly improved performance to baseline levels (Dice score 0.77, P < .001), with high correlation (Spearman r = 0.98) in total lesion volumes.

Conclusions:

• A modest addition of relevant, heterogeneous training data from an external institution can successfully adapt a pre-trained brain MRI segmentation model for use at that new site.
• The findings highlight the importance of data diversity over sheer volume when augmenting training sets for improved cross-institutional performance.
• This approach facilitates the reliable application of AI models in diverse clinical settings, enhancing diagnostic capabilities.