Best Practices for Advancing Neuroimaging Tools on The Edge
Poster No:
1894
Submission Type:
Abstract Submission
Authors:
Mohamed Masoud1, Pratyush Reddy1, Farfalla Hu1, Sergey Plis1
Institutions:
1Georgia State University, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
Processing volumetric medical images within web browsers poses unprecedented challenges emerging from the inherent constraints of browser environments. This abstract outlines Brainchop's (https://github.com/neuroneural/brainchop) processing pipeline, an innovative in-browser neuroimaging tool, and evaluates the performance of models across diverse software and hardware configurations, providing best practices for creating client-side neuroimaging tools and analysis of its performance from the lens of Causality. The outcomes highlight the practical viability of client-side processing for volumetric data attributable to the robust MeshNet architecture.
Methods:
Brainchop[1] performs MRI segmentation using Meshnet models [2-3] trained in PyTorch with the Human Connectome [4] dataset and converted to tensorflow.js [5] for in-browser inference. It offers a range of volumetric segmentation tasks (Fig. 1a). Table 1 provides a list of the models along with their performance. We analyzed anonymously collected telemetry data to evaluate the tool's performance and identify factors that affect its success rate. Data preprocessing involved cleaning by excluding extreme outliers and removing features with high correlation coefficients (Threshold > 0.95). The selected dataset is devoid of missing values, and categorical data is encoded using a label encoder. For regression models, one-hot encoding is employed to capture the effect of each categorical variable independently. Statistical analysis utilized a 95% confidence interval for null hypothesis testing (P<0.05).
Brainchop's performance is improved through interventions like patching and cropping applied to input data. Causal analysis to accurately estimate their effects involves treating each intervention as a distinct treatment and isolating its effect from potential confounders. To isolate each effect, the Inverse Probability of Treatment Weighting (IPTW) [6] is used. The Average Treatment Effect (ATE) can be estimated as::
ATE= p( Outcome = 1 ∣ do (Treatment = 1) ) − p( Outcome = 1 ∣ do( Treatment = 0 ))
This represents the probability of success rate when applying the treatment (e.g., patching or cropping) versus when not applying it.
Brainchop's performance is improved through interventions like patching and cropping applied to input data. Causal analysis to accurately estimate their effects involves treating each intervention as a distinct treatment and isolating its effect from potential confounders. To isolate each effect, the Inverse Probability of Treatment Weighting (IPTW) [6] is used. The Average Treatment Effect (ATE) can be estimated as::
ATE= p( Outcome = 1 ∣ do (Treatment = 1) ) − p( Outcome = 1 ∣ do( Treatment = 0 ))
This represents the probability of success rate when applying the treatment (e.g., patching or cropping) versus when not applying it.
Results:
Brainchop achieved an 82% success rate from May 2022 to May 2023 despite diverse user-side computational resources. A statistical analysis of the telemetry data using the Chi-square test revealed an adequate sample size with a power of 0.963 (α = 0.05). We incorporated full-volume and sub-volume (Failsafe) models to address edge computational limitations. Our analysis showed that failed statuses were primarily linked to GPU memory limitations. While sub-volume models had higher success rates, they also incurred slower inference times, reduced accuracy, and merge overheads compared to full-volume.
Patching treatment, affected by cropping, had an independent effect on the success rate. Estimation of patching effect using IPTW showed a 6.23% success rate increase, a 24.31 second inference time increase, and minimal post processing time change. Our Chi-square tests confirmed a significant correlation (p-value 2-09) between cropping and success rates in full volume inference, with a 99.9% statistical power. IPTW estimates for the cropping effect indicate an 18.12% success rate increase, a 5.26-second decrease in inference time, and a 6.83-second reduction in post processing time. Finally, Fig. 1- b and c depict the interdependence between success rates over time, with GPU and model choice influencing each other.
Patching treatment, affected by cropping, had an independent effect on the success rate. Estimation of patching effect using IPTW showed a 6.23% success rate increase, a 24.31 second inference time increase, and minimal post processing time change. Our Chi-square tests confirmed a significant correlation (p-value 2-09) between cropping and success rates in full volume inference, with a 99.9% statistical power. IPTW estimates for the cropping effect indicate an 18.12% success rate increase, a 5.26-second decrease in inference time, and a 6.83-second reduction in post processing time. Finally, Fig. 1- b and c depict the interdependence between success rates over time, with GPU and model choice influencing each other.
Conclusions:
Our research has identified a statistically significant correlation between the patching and cropping techniques and both the temporal aspects and the success rate of Brainchop. Noteworthy is the observation that Brainchop has demonstrated a high success rate. Our findings highlight the imperative to refine cropping techniques for optimal outcomes and show a potential to reveal more insights that can improve Brainchop's functionality.
Modeling and Analysis Methods:
Methods Development 1
Segmentation and Parcellation
Neuroinformatics and Data Sharing:
Workflows 2
Informatics Other
Keywords:
Data analysis
Design and Analysis
Informatics
Machine Learning
MRI
Open-Source Software
Segmentation
Statistical Methods
1|2Indicates the priority used for review
Provide references using author date format
[2] Fedorov, A., Johnson, J., Damaraju, E., Ozerin, A., Calhoun, V., & Plis, S. (2017), “End-to-end learning of brain tissue segmentation from imperfect labeling”, IEEE International Joint Conference on Neural Networks (IJCNN). https://doi.org/10.1109/IJCNN.2017.7966333
[3] Yu, F., Koltun, V. (2016), “Multi-scale context aggregation by dilated convolutions”, arXiv.https://doi.org/10.48550/arXiv.1511.07122
[4] D. C. Van Essen, S. M. Smith, D. M. Barch, T. E. Behrens, E. Yacoub, K. Ugurbil, W.-M. H. Consortium et al., (2013), “The wu-minn human connectome project: an overview”, Neuroimage, vol. 80, pp. 62–79.
[5] Smilkov, D., Thorat, N., Assogba, Y., Yuan, A., Kreeger, N., & et, al. (2019), “TensorFlow.js: Machine learning for the web and beyond”, arXiv. https://doi.org/10.48550/arXiv.1901.05350
[6] N. Chesnaye, V. Stel, G. Tripepi, F. Dekker, E. Fu, C. Zoccali, K. Jager (2022), An introduction to inverse probability of treatment weighting in observational research , Clinical Kidney Journal