Traumatic Brain Injury Segmentation using an Ensemble of Encoder-decoder Models

2025-09-29

AI Summary

Moderate-to-severe traumatic brain injury (TBI) lesions exhibit substantial inter-subject heterogeneity in size, number, and spatial distribution on MRI, severely compromising the accuracy of downstream tasks such as image registration and brain parcellation. To address this challenge, we propose a multi-encoder–decoder ensemble method built upon the nnUNet framework, integrating diverse backbone architectures—including ResNet and Swin Transformer—and incorporating an adaptive post-processing strategy to enhance segmentation robustness and precision for complex TBI lesions. Evaluated on the T1-weighted MRI dataset from the AIMS-TBI 2025 Challenge, our method achieves an overall Dice score of 0.5973 (0.8514 for lesion-free and 0.4711 for lesion-bearing subjects), with an accuracy of 0.8451, ranking among the top six submissions. This work establishes a highly reliable technical pipeline for automated quantitative neuroimaging analysis of TBI.

Technology Category

• Machine Learning: Ensemble Methods
• Computer Vision: Segmentation
• Humans and AI: Brain-Sensing and Analysis

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

The identification and segmentation of moderate-severe traumatic brain injury (TBI) lesions pose a significant challenge in neuroimaging. This difficulty arises from the extreme heterogeneity of these lesions, which vary in size, number, and laterality, thereby complicating downstream image processing tasks such as image registration and brain parcellation, reducing the analytical accuracy. Thus, developing methods for highly accurate segmentation of TBI lesions is essential for reliable neuroimaging analysis. This study aims to develop an effective automated segmentation pipeline to automatically detect and segment TBI lesions in T1-weighted MRI scans. We evaluate multiple approaches to achieve accurate segmentation of the TBI lesions. The core of our pipeline leverages various architectures within the nnUNet framework for initial segmentation, complemented by post-processing strategies to enhance evaluation metrics. Our final submission to the challenge achieved an accuracy of 0.8451, Dice score values of 0.4711 and 0.8514 for images with and without visible lesions, respectively, with an overall Dice score of 0.5973, ranking among the top-6 methods in the AIMS-TBI 2025 challenge. The Python implementation of our pipeline is publicly available.

Problem

• Automated segmentation of traumatic brain injury lesions
• Addressing lesion heterogeneity in size and location
• Improving accuracy for neuroimaging analysis tasks

Innovation

• Uses ensemble of encoder-decoder models
• Leverages nnUNet framework for segmentation
• Implements post-processing to enhance metrics