Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Automated Lesion Segmentation Using Acute and Chronic Stroke MRI with nnU-Net
Poster Session C, Friday, October 25, 4:30 - 6:00 pm, Great Hall 3 and 4
Tammar Truzman1, Ajay Halai1, Matt Lambon-Ralph1; 1University of Cambridge
Introduction : Stroke is a leading cause of language deficits (aphasia) and cognitive-clinical neuroscience research has explored the relationship between brain, behaviour and recovery after stroke for many decades. Lesion segmentation is central for both clinical and research purposes, particularly for lesion-symptom mapping1,2 and estimating functional and structural disconnections3. Therefore, accurate and reliable lesion segmentation is essential, with manual tracing considered to be the gold standard. However, manual tracing is time-consuming, subject to inter-tracer variability and unfeasible for large/multi-site studies. Many automated algorithms have been prosed (i.e., 3–5); however, their performance is only moderate for multi-site datasets6 (max DICE=0.55 compared to manual tracing inter-rater reliability DICE=0.73). Additional significant challenges arise in longitudinal studies, where a ‘lesioned’ voxel/region may change over time making it difficult to know which areas should be marked as lesioned given different imaging modalities, and questioning the validity of using acute lesion tracing for chronic data. In this study, we took large multisite datasets in acute and chronic stroke and apply nnU-Net7 for lesion segmentation. We expected that: (1) nnU-Net models would outperform previous benchmark algorithms; and (2) perfusion MRI would be crucial for lesion identification in the acute phase, whereas structural T1 would be primary for lesion mapping in chronic stroke. Methods: We collated open-source datasets of acute8,9 and chronic stroke10,11 along with data from the MRC Cognition and Brain Sciences Unit and Korea University (KU), which had MRI and manual tracing. The acute dataset included N=1798 cases based on DWI b=1000, DWI apparent diffusion coefficient, FLAIR and T1, while the chronic dataset included N=1171 cases based on T1 only. Separate nnU-Net models were built for acute and chronic datasets using the default settings in nnU-Net (v2). Model performance (DICE coefficient) was tested using 5-fold cross validation with all data except the KU dataset, which was used for external out-of-sample validation. As we had multi-model MRI for acute data, we tested which scan types and combinations were most important for accurate lesion estimation. Results: The nnU-Net chronic model surpassed the accuracy of previously used algorithms with a DICE=0.68; importantly nearing manual tracing performance. The results for the acute data showed that: using T1 only leads to poor performance as expected (DICE=0.35), but FLAIR was also ineffective (DICE=0.47). In contrast, using DWI b=1000 scans led to high performance (DICE=0.74) and neared the reliability of manual tracing12 (DICE=0.76 and 0.79 for inter- and intra-rater evaluations, respectively). Conclusion: In this study, we were able to demonstrate that the nnU-Net framework for lesion estimation can outperform previous benchmarks and approaches the reliability for manual tracing for both acute and chronic stroke. This would aid large scale studies and also guide which imaging sequence(s) offer the best lesion segmentation in the acute stage. We will make the models open source with accompanying code, which can be tested further or refined with more data or better algorithms. References: Not enough space
Topic Areas: Methods, Computational Approaches