Related Experiment Video
Updated: Oct 2, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
48.5K
Improving interobserver agreement and performance of deep learning models for segmenting acute ischemic stroke by
Chun-Jung Juan1,2,3,4,5, Shao-Chieh Lin2,6, Ya-Hui Li2,7
1Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan, Republic of China.
European Radiology
|February 24, 2022
Summary
Optimizing acute ischemic stroke lesion segmentation using deep learning models requires combining diffusion-weighted imaging (DWI) with an apparent diffusion coefficient (ADC) threshold of 0.6 × 10⁻³ mm²/s. This approach significantly improves segmentation accuracy and reduces differences between observers and models.
Area of Science:
- Radiology
- Medical Imaging
- Artificial Intelligence in Medicine
Background:
- Accurate segmentation of acute ischemic stroke (AIS) lesions is crucial for diagnosis and treatment planning.
- Deep learning models (DLMs) show promise for automating this process, but their performance can be influenced by imaging parameters.
- The role of apparent diffusion coefficient (ADC) thresholds in conjunction with diffusion-weighted imaging (DWI) for DLM segmentation of AIS needs further investigation.
Purpose of the Study:
- To evaluate the impact of different ADC thresholds on the agreement among human observers and DLMs for AIS segmentation.
- To assess the segmentation performance of DLMs using various ADC thresholds.
- To determine the optimal ADC threshold for improving DLM accuracy in AIS lesion detection.
Main Methods:
- Twelve DLMs were trained on DWI-ADC data from 76 AIS patients using six different ADC thresholds.
- Model performance was validated on independent datasets from two hospitals (67 and 78 patients).
- Agreement was assessed using Bland-Altman plots and intraclass correlation coefficients (ICC); segmentation accuracy was measured by Dice similarity coefficient (DSC).
Main Results:
- Excellent interobserver and intraobserver agreement were achieved for manual segmentation (ICC > 0.98).
- Combining DWI with an ADC threshold of 0.6 × 10⁻³ mm²/s significantly reduced the limit of agreement (0.59 cm²) compared to DWI alone (11.23 cm²).
- DLM segmentation performance improved substantially, with DSC increasing from 0.738 (DWI alone) to 0.971 (DWI with ADC threshold of 0.6 × 10⁻³ mm²/s).
Conclusions:
- An ADC threshold of 0.6 × 10⁻³ mm²/s, when combined with DWI, enhances DLM segmentation of AIS lesions.
- This optimized approach reduces interobserver and inter-DLM variability, leading to more consistent and accurate lesion identification.
- The findings suggest that incorporating specific ADC thresholds is vital for maximizing the efficacy of DLMs in stroke imaging analysis.
Keywords:
Apparent diffusion coefficientDeep learningDice similarity coefficientDiffusion magnetic resonance imagingIschemic stroke
