State-of-the-Art Stroke Lesion Segmentation at 1/1000th of Parameters
Poster No:
1650
Submission Type:
Abstract Submission
Authors:
Alex Fedorov1, Yutong Bu1, Sarah Wilson2, Leonardo Bonilha2, Roger Newman-Norlund2, Xiao Hu1, Chris Rorden2, Sergey Plis3
Institutions:
1Emory University, Atlanta, GA, 2University of South Carolina, Columbia, SC, 3Georgia State University, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
Accurate segmentation of brain lesions in MRI can aid in diagnosis, treatment planning, and monitoring of neurological conditions like chronic stroke. While manual lesion delineation is often subjective and time-consuming [1], automatic segmentation offers a more rapid and objective solution. The state-of-the-art architectures such as MedNeXt [2] and U-MAMBA [3], based on the U-Net [4], deliver high accuracy but require millions of parameters and intricate encoder-decoder frameworks. Yet these advances come at the cost of increased complexity or computational demands. Such large models are challenging to deploy in resource-limited or edge-based environments and may necessitate sending sensitive patient data to remote servers. In contrast, our proposed architecture achieves state-of-the-art stroke lesion segmentation with as little as 1/1000th the parameters of comparable models, enabling on-edge inference [5] that preserves data privacy, maintains accuracy, and removes the need for extensive computational resources.
Methods:
We revisit MeshNet [6] and adapt it for whole-brain chronic lesion segmentation on full-sized 256×256×256 MRI volumes without subvolume sampling or 2D slices. The revised MeshNet employs a novel dilation pattern inspired by U-Net-like encoder-decoder paradigms. Dilation rates initially increase (1, 2, 4, 8, 16) to capture broad contextual features, then decrease (16, 8, 4, 2, 1) to regain fine-grained detail without explicit down/upsampling or skip connections. Zero padding was applied to maintain the input size across all layers with X channels. We trained and evaluated MeshNet-X variants (5,682 to 147,474 parameters) and compared them against common segmentation architectures that typically require millions of parameters. We optimized a weighted cross-entropy loss with label smoothing using AdamW with OneCycleLR.
We performed experiments on the Aphasia Recovery Cohort (ARC) [7] dataset, comprising T2-weighted MRI scans from 230 individuals with chronic stroke lesions. Preprocessing involved using FreeSurfer's mri_convert conform function to resample images to a uniform 1mm isotropic resolution and scale them to a 256×256×256 with intensities normalized to the 0–1 range. Nested cross-validation was employed, with three outer and three inner folds per outer fold.
We performed experiments on the Aphasia Recovery Cohort (ARC) [7] dataset, comprising T2-weighted MRI scans from 230 individuals with chronic stroke lesions. Preprocessing involved using FreeSurfer's mri_convert conform function to resample images to a uniform 1mm isotropic resolution and scale them to a 256×256×256 with intensities normalized to the 0–1 range. Nested cross-validation was employed, with three outer and three inner folds per outer fold.
Results:
MeshNet-26 (Fig. 1, Tab. 1) achieved a median DICE of 0.876 with 147,474 parameters-approximately 50× fewer than U-MAMBA-BOT (0.870 DICE) and 120× fewer than MedNeXt-M (0.868 DICE). MeshNet-16, with 56,194 parameters, attained a DICE of 0.873, offering 130× and 310× less parameters. Even the minimal MeshNet-5, at 5,682 parameters, reached 0.848 DICE (just under the 0.85 threshold [8]) with 1,300× and 3,089× fewer parameters, demonstrating robust performance at a fraction of the complexity. Plotting inverse parameter count against median DICE (Fig. 2) placed MeshNet on the Pareto frontier, confirming that tens of millions of parameters are not required for accurate whole-brain segmentation. Visual evaluations (Fig. 3) showed that MeshNet-16 and -26 closely aligned with lesion boundaries and captured subtle edges.
Conclusions:
This work demonstrates the feasibility of achieving state-of-the-art stroke lesion segmentation with only a fraction of the parameters required by state-of-the-art architectures. By carefully engineering a dilation pattern and performing whole-brain inference without down/upsampling or skip connections, our adapted MeshNet achieves robust accuracy while dramatically reducing model complexity. Such parameter efficiency makes the approach especially suitable for resource-constrained settings, web-based applications, and real-time clinical workflows where computational resources and privacy concerns are paramount. Future research will evaluate MeshNet across broader datasets and further refine its design, ensuring that its remarkable balance of accuracy, speed, and compactness continues to drive accessible, high-quality medical image analysis.
Modeling and Analysis Methods:
Methods Development 2
Segmentation and Parcellation 1
Keywords:
Machine Learning
Segmentation
STRUCTURAL MRI
Other - stroke
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
[2] Roy, S., Koehler, G., Ulrich, C., Baumgartner, M., Petersen, J., Isensee, F., Jaeger, P. F., & Maier-Hein, K. H. (2023). MedNeXt: Transformer-driven scaling of ConvNets for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 405–415). Springer.
[3] Ma, J., Li, F., & Wang, B. (2024). U-MAMBA: Enhancing long-range dependency for biomedical image segmentation. arXiv Preprint, arXiv:2401.04722. https://arxiv.org/abs/2401.04722
[4] Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. In N. Navab, J. Hornegger, W. M. Wells, & A. F. Frangi (Eds.), Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Part III (pp. 234–241). Springer.
[5] Plis, S. M., Masoud, M., Hu, F., Hanayik, T., Ghosh, S. S., Drake, C., ... & Rorden, C. (2024). Brainchop: Providing an Edge Ecosystem for Deployment of Neuroimaging Artificial Intelligence Models. Aperture neuro, 4.
[6] Fedorov, A., Johnson, J., Damaraju, E., Ozerin, A., Calhoun, V., & Plis, S. (2017). End-to-end learning of brain tissue segmentation from imperfect labeling. In 2017 International Joint Conference on Neural Networks (IJCNN) (pp. 3785–3792). IEEE.
[7] Gibson, M., Newman-Norlund, R., Bonilha, L., Fridriksson, J., Hickok, G., Hillis, A. E., den Ouden, D.-B., & Rorden, C. (2024). The Aphasia Recovery Cohort, an open-source chronic stroke repository. Scientific Data, 11(1), 981.
[8] Liew, S.-L., Lo, B. P., Donnelly, M. R., Zavaliangos-Petropulu, A., Jeong, J. N., Barisano, G., Hutton, A., Simon, J. P., Juliano, J. M., Suri, A., … (2022). A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1), 320.