Source-linked AI summary
Deep Learning Segmentation of Diffusion-Weighted MRI Acute Ischaemic Stroke: A Pragmatic Evaluation Across Three Datasets
Atle Bjørnerud, Till Schellhorn, Thor H. Skattør, Terje Nome, Jon André Ottesen, Anne Hege Aamodt, Bradley J MacIntosh
TL;DR
Routine clinical translation of DWI segmentation models remains limited, alongside the need for high-quality AIS segmentation. This study found that a single-input, unpreprocessed nnU-Net enabled fast and accurate segmentation and may facilitate efficient use in practice.
Problem
Routine clinical translation of DWI segmentation models remains limited, while high-quality AIS segmentation remains an ongoing image-analysis challenge.
Method
The study evaluated nnU-Net models using DWI segmentation with minimal preprocessing, including brain extraction comparisons and alternative input configurations.
Results
Median DSC was 0.84 for the base model, which compared favourably across all performance test metrics.
Takeaways & Limitations
A single-DWI nnU-Net without preprocessing enabled fast and accurate AIS lesion segmentation and may facilitate efficient inference.
Takeaways & Limitations
EPI fundamentally suffers from spatial and intensity limitations, and the study reports additional limitations that should be discussed.
Abstract
from arXiv · showhide
Objective: Diffusion-weighted MRI (DWI-MRI) is the gold standard for visualizing and quantifying acute ischaemic stroke (AIS). Although deep learning methods can accurately segment AIS lesions, the optimal image inputs and model architecture remain uncertain. We evaluated whether accurate AIS lesion segmentation can be achieved using a pragmatic deep learning approach with minimal preprocessing and clinically feasible inference times. Materials and Methods: Self-configured nnU-Net models were trained on 1,744 DWI cases from local, national, and open-access datasets and tested on 436 cases. Four experimental conditions were evaluated using five-fold cross-validation: with or without brain extraction and using either DWI alone or DWI plus apparent diffusion coefficient (ADC) images as inputs. Two architectures were compared: the baseline nnU-Net (base) and a residual encoder nnU-Net (ResEnc). Performance was benchmarked against the DeepISLES ensemble model from the 2022 ISLES challenge. Results: In the test set (n=436), the base model achieved a median (IQR) Dice similarity coefficient (DSC) of 0.84 (0.19). For the base model, only two of six pairwise comparisons between input configurations showed significant differences. ResEnc produced small but significant improvements in DSC compared with the base model for DWI, DWI+brain extraction, and DWI+ADC inputs (all p<0.02), but not for DWI+ADC+brain extraction (p>0.50). The base model significantly outperformed DeepISLES, particularly in patients with smaller infarct volumes (signed-rank test, p<0.01). Conclusions: A baseline nnU-Net trained on DWI alone, without preprocessing, enabled fast and accurate AIS lesion segmentation. This streamlined approach may facilitate clinical research and support acute stroke imaging workflows
Introduction
AIS lesion segmentation on DWI is clinically promising but remains difficult to translate into routine radiology because existing workflows can require complex preprocessing and multiple image inputs. This study evaluates whether a simplified nnU-Net approach using routinely acquired DWI can provide accurate segmentation with minimal preprocessing and rapid inference.
- Clinical motivation: DWI-MRI is used to visualize acute ischaemic stroke, while deep learning is being investigated for automated AIS lesion delineation.Prior studies and international challenges report strong performance across heterogeneous datasets.
- Clinical motivation: Clinical translation of DWI segmentation models remains limited partly because of methodological complexity and workflow-related challenges.The stated goal is robust and accurate automated lesion delineation that can support quantitative imaging, prognostication, and treatment-effect evaluation.
- Practical constraints: High-quality AIS segmentation is sought with as few DWI inputs and preprocessing steps as possible.Common approaches use brain extraction and/or additional ADC maps, increasing imaging-pipeline complexity.
- Practical constraints: Brain extraction is difficult to verify routinely because visual inspection is impractical in resource-limited clinical settings.Its performance may also vary on DWI because of poorer image quality, distortions, susceptibility artefacts, and limited representative training data.
- Practical constraints: Using both DWI and ADC complicates deployment because ADC availability and series organization vary across MRI vendors and hospital PACS systems.These variations make automated input handling non-trivial and can introduce sources of error that reduce robustness in real-world environments.
- Study aim: The study evaluates a pragmatic nnU-Net solution that avoids image alignment, uses minimal preprocessing, and aims for rapid inference directly from routinely acquired DWI.The authors propose that this could lower barriers to implementation and improve integration into radiological workflows.
Material and Methods
The study retrospectively assembled annotated acute-stroke MRI from three cohorts and evaluated multiple nnU-Net input and architecture configurations against DeepISLES. Performance was assessed with established lesion-overlap, lesion-wise, volume, and count metrics using cross-validation and non-parametric comparisons.
- Datasets and annotations: MRI data came from Nor-COAST, OSCAR, and the open-access SOOP dataset, with DWI, derived ADC, and other clinical MRI data represented.AIS lesions were annotated on high b-value DWI, and only acute lesions were included from SOOP.
- Datasets and annotations: Cases were excluded for intracerebral haemorrhage, excessive DWI motion, missing ground-truth masks, or corrupted data.Ground-truth segmentations were produced by experienced neuroradiologists or trained neuroscientists using semi-automatic or manual tools.
- Model configurations: Four input conditions combined DWI alone or DWI plus ADC with or without brain extraction.The resulting configurations were evaluated using a standard base nnU-Net and a residual-encoder ResEnc architecture.
- Model configurations: ResEnc required approximately 3 times more GPU memory and approximately 4 times longer runtime than the base model.Both model types used default auto-configurations apart from reducing training from 1000 to 500 epochs.
- Evaluation: All nnU-Net configurations used five-fold cross-validation, with the five inference models ensembled by averaging softmax predictions.Training and prediction compute times were recorded, and models were compared with the externally validated DeepISLES ensemble.
- Evaluation: Performance was evaluated using DSC, lesion-wise F1-score, absolute volume difference, and absolute lesion count difference.Non-parametric tests compared configurations, datasets, and nnU-Net with DeepISLES, including analyses by ground-truth lesion-volume range.
Results
Across three datasets, the baseline nnU-Net achieved strong lesion-segmentation performance with DWI alone and no brain extraction, while ResEnc offered only modest configuration-dependent gains. Compared with DeepISLES, the base model performed at least as well overall, with advantages for smaller lesions and faster inference.
- Datasets: 2,180 AIS cases remained after exclusions, with 1,744 used for training/validation and 436 for testing.Twelve OSCAR test cases were excluded from the DeepISLES comparison because FLAIR data were missing.
- nnU-Net performance: 0.84 (0.19) was the base model’s median (IQR) DSC using DWI alone without brain extraction.The same configuration achieved a lesion-wise F1-score of 0.77 (0.43), AVD of 1.37 (3.96) mL, and ALD of 1 (3).
- Input and architecture comparisons: No significant DSC difference was found between DWI-only and DWI+ADC inputs, while ResEnc improved DSC over the base model for selected configurations.ResEnc gains were observed for DWI, DWI+brain extraction, and DWI+ADC, but not for DWI+ADC+brain extraction.
- Efficiency: 13 hours versus 55 hours were required for one-fold training of the base and ResEnc models, with inference times of approximately 5.0 versus 11.0 seconds per case.The base model was selected for comparison with DeepISLES because it was faster and had comparable performance relative to ResEnc.
- Comparison with DeepISLES: 0.84 (0.19) versus 0.81 (0.19) was the median (IQR) DSC for nnU-Net versus DeepISLES across the three test sets, favoring nnU-Net.The difference was statistically significant, and nnU-Net showed non-inferior or improved performance across all reported metrics.
Discussion
The discussion supports a streamlined DWI-only nnU-Net approach while identifying modest architecture gains, input trade-offs, benchmark advantages, and limitations affecting broader deployment.
- Architecture: The ResEnc model produced only a marginal Dice improvement over the base model, while incurring greater compute cost and longer inference times.The discussion frames the larger architecture’s gains against its implementation burden.
- Benchmarking: The base nnU-Net was not inferior to DeepISLES and significantly outperformed it under certain experimental conditions.The comparison favored nnU-Net particularly for smaller lesions and in the Nor-COAST cohort.
- Preprocessing: Skull-stripping did not meaningfully improve segmentation and may introduce inaccuracies, additional analysis steps, and clinical implementation issues.DWI signal dropout can complicate brain extraction, and correcting it may require scans that are not clinically routine.
- Input channels: Adding ADC tended to protect against false positives caused by T2 shine-through, including a correctly segmented case among 436 test cases.The discussion presents ADC as potentially useful for distinguishing high DWI signal from restricted diffusion, despite limited overall performance gains.
- Generalization: DeepISLES performed worse on smaller lesions and failed to detect lesions in seven of 51 Nor-COAST cases, whereas nnU-Net failed in one.The authors suggest that Nor-COAST image quality or contrast may have differed substantially from DeepISLES training data.
- Limitations: The study’s scope prioritized sufficient performance on regional institutional data over robustness across the widest range of image qualities and protocols.Unexpected incidental or co-pathologies also produced segmentation errors, limiting confidence outside the intended clinical setting.