Indirect Volume Estimation for Acute Ischemic Stroke from Diffusion Weighted Image Using Slice Image Segmentation

Seung-Ah Lee, Jae‐Won Jang, Sang Won Park, Pum-Jun Kim, Na-Young Yeo, Chulho Kim, Yoon Kim, Hyun-Soo Choi, Seongheon KimJournal of Personalized Medicine 2022 · DOI

TL;DR

This paper introduces a novel indirect volume estimation method for acute ischemic stroke (AIS) using diffusion-weighted imaging (DWI). By converting 3D DWI scans into 2D slices and employing a pretrained segmentation model, their approach achieves more accurate and reliable AIS volume estimation compared to direct 3D segmentation, offering a valuable tool for clinical decision-making.

What Problem Does This Paper Address?

Acute ischemic stroke (AIS) is a leading cause of disability, and timely intervention is critical for improving patient outcomes. Accurately assessing the volume of the infarct core on diffusion-weighted imaging (DWI) is crucial for guiding treatment decisions, such as reperfusion therapies. However, manual segmentation of AIS lesions on DWI is a laborious, time-consuming process that can be subjective and prone to errors. Existing automated segmentation methods often struggle with the complexity of AIS, including its varied subtypes, artifacts, and ill-defined boundaries, and can be computationally intensive. This paper aims to bridge this gap by proposing an objective, fast, and accurate method for estimating AIS volume directly from DWI.

Methods

The researchers developed two primary approaches for AIS volume estimation:

  • Direct Volume Estimation (3D Segmentation): This method utilizes a 3D U-Net architecture, a convolutional neural network (CNN) designed for processing volumetric data. It employs a symmetric encoder-decoder structure with skip connections to capture both contextual and spatial information. To address the challenges of limited data and class imbalance in AIS segmentation, they utilized transfer learning with pretrained models for auto implants and tumors, and employed patch-based segmentation.
  • Indirect Volume Estimation (2D Segmentation): This novel approach converts the 3D DWI volume into a series of 2D slice images. Each slice is then processed by a 2D U-Net model. This strategy increases the amount of data available for training and allows the model to capture broader brain cross-sectional information, potentially mitigating hierarchical imbalances. Crucially, they used a pretrained model for brain glioma segmentation, which was found to be more similar to AIS segmentation than models for auto implants or tumors. Volume is then estimated by aggregating the segmented pixels across slices, taking into account pixel volume derived from image spacing and slice thickness.

Both methods were trained using a Dice loss function, and their performance was evaluated using standard segmentation metrics (sensitivity, specificity, F1-score, Jaccard index) and volume estimation metrics (Volume Similarity - VS, Mean Absolute Error - MAE).

Datasets

The study utilized a retrospective dataset of DWI scans from 2159 participants with AIS, collected from two hospitals: Hallym University Chuncheon Sacred Heart Hospital (HUCSHH) and Kangwon National University Hospital (KNUH). This dataset was split into an internal validation set (1179 participants) and an external validation set (980 participants). Additionally, 121 DWI images from control subjects were used to evaluate the segmentation algorithm's performance on non-AIS cases.

Key Insights and Ideas

  • 2D Segmentation Outperforms 3D for AIS: Surprisingly, the indirect 2D segmentation approach, despite being conceptually simpler, demonstrated superior performance in both segmentation and volume estimation compared to the direct 3D segmentation method. This suggests that for AIS, which can be small and sparsely distributed, processing slice-by-slice with appropriate context (e.g., using adjacent slices as input channels) might be more effective than a purely volumetric approach.
  • Pretrained Models are Key: The use of pretrained models, particularly for the indirect 2D segmentation using a glioma segmentation model, significantly boosted performance. This highlights the power of transfer learning in medical image analysis, especially when dealing with limited or imbalanced datasets. The similarity in task (lesion segmentation) and input modality (DWI) between glioma and AIS likely contributed to this success.
  • Indirect Estimation is More Reliable for Volume: While segmentation metrics are important, the ultimate goal is accurate volume estimation. The indirect method showed significantly higher Volume Similarity (VS) and lower Mean Absolute Error (MAE) for AIS volume, indicating a more reliable and accurate estimation of the infarct size.

Key Results

  • Segmentation Performance: The pretrained indirect model achieved the highest F1-scores: 76.02% for internal validation and 77.23% for external validation. The Jaccard indices were also impressive at 62.12% (internal) and 63.82% (external). These figures were notably higher than those achieved by the direct segmentation models.
  • Volume Estimation Reliability: The indirect volume estimation with a pretrained glioma model demonstrated excellent results: 93.3% VS and 0.797 cc MAE for internal validation, and 89.2% VS and 2.468 cc MAE for external validation. These metrics indicate a strong agreement between the predicted and actual AIS volumes.
  • FP Error Reduction: The best model (indirect pretrained) showed minimal false positive (FP) errors when validated on a normal group, with MAE values of 0.028 cc (internal) and 0.009 cc (external), suggesting high specificity in identifying actual lesions.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that segmenting sparse and small-sized AIS can still be challenging for the indirect method, as these cases may not provide enough axial slices for effective learning.
  • The reliance on DWI from MRI, which is typically available in larger medical centers, is noted as a limitation.
  • Future Work:
    • The authors plan to focus on improving the performance of direct volume estimation for small AIS cases.
  • They aim to develop a system that can predict the patient's modified Rankin Score (mRS), a measure of disability, by integrating image information with clinical patient data.
  • Future studies could explore validation with other imaging modalities like angiography or computed tomography to further reduce "door-to-needle" time.

Summary updated

Abstract

The accurate estimation of acute ischemic stroke (AIS) using diffusion-weighted imaging (DWI) is crucial for assessing patients and guiding treatment options. This study aimed to propose a method that estimates AIS volume in DWI objectively, quickly, and accurately. We used a dataset of DWI with AIS, including 2159 participants (1179 for internal validation and 980 for external validation) with various types of AIS. We constructed algorithms using 3D segmentation (direct estimation) and 2D segmentation (indirect estimation) and compared their performances with those annotated by neurologists. The proposed pretrained indirect model demonstrated higher segmentation performance than the direct model, with a sensitivity, specificity, F1-score, and Jaccard index of 75.0%, 77.9%, 76.0, and 62.1%, respectively, for internal validation, and 72.8%, 84.3%, 77.2, and 63.8%, respectively, for external validation. Volume estimation was more reliable for the indirect model, with 93.3% volume similarity (VS), 0.797 mean absolute error (MAE) for internal validation, VS of 89.2% and a MAE of 2.5% for external validation. These results suggest that the indirect model using 2D segmentation developed in this study can provide an accurate estimation of volume from DWI of AIS and may serve as a supporting tool to help physicians make crucial clinical decisions.
BibTeX
@article{lee2022indirect,
  author = {Seung-Ah Lee and Jae‐Won Jang and Sang Won Park and Pum-Jun Kim and Na-Young Yeo and Chulho Kim and Yoon Kim and Hyun-Soo Choi and Seongheon Kim},
  title = {{Indirect Volume Estimation for Acute Ischemic Stroke from Diffusion Weighted Image Using Slice Image Segmentation}},
  journal = {Journal of Personalized Medicine},
  year = {2022},
  doi = {10.3390/jpm12040521},
}
Figure