Cerebral Small Vessel Disease Biomarkers Detection on MRI-Sensor-Based Image and Deep Learning

Yi‐Zeng Hsieh, Yu Luo, Chen Pan, Mu Su, Chi Jen Chen, Kevin Li‐Chun HsiehSensors 2019 · DOI

TL;DR

This paper presents a deep learning approach using Convolutional Neural Networks (CNNs) to automatically detect cerebral small vessel disease (CSVD) from MRI scans. The system aims to assist clinicians in quickly identifying and quantifying these lesions, which are a significant cause of stroke, thereby improving diagnostic speed and accuracy.

What Problem Does This Paper Address?

Cerebrovascular disease is a leading cause of death globally and in Taiwan, with stroke being a primary manifestation. Cerebral small vessel disease (CSVD) is a common and often fatal cause of stroke, characterized by lesions in small blood vessels within the brain. These lesions are typically small (less than 1.5 cm) and can be difficult for clinicians to detect and precisely locate on MRI scans, requiring extensive examination time. The paper addresses the need for an automated, computer-aided diagnosis (CAD) system that can efficiently identify and characterize these subtle lesions, helping to determine the probability and severity of stroke.

Methods

The core of the proposed method is a Convolutional Neural Network (CNN) architecture designed for image recognition and classification. The authors developed a custom 7-layer CNN, optimized for CPU execution to minimize computational resource requirements.

The process involves several key steps:

  1. Data Preprocessing: This stage focuses on isolating the brain region from the MRI scan. It includes image binarization, removal of the cranial shell (head bone) by identifying and removing the largest connected component, and inverse binarization to obtain the cerebrum. Median filtering is then applied to reduce noise.
  2. Image Segmentation (Patch-based CNN): To handle the small size and varied appearance of CSVD lesions, the MRI images are segmented into smaller patches (7x7 pixels). This patch-based approach allows the CNN to focus on local features and achieve better detection results.
  3. CNN Model Architecture: A 7-layer CNN is employed, consisting of convolutional layers, max-pooling layers, and fully connected layers, culminating in a softmax classifier. The network is trained to classify the patches into four categories: non-brain block, normal brain, normal brain position, and cerebrovascular disease block (lesion).
  4. Training and Evaluation: The model is trained using a dataset of annotated MRI images. The authors employ a 10-fold cross-validation strategy to mitigate overfitting. The performance is evaluated using metrics like accuracy, F1-score, precision, and recall, and compared against other models.

Datasets

The study utilized a dataset comprising MRI images from 50 patients provided by Taipei Medical University-Shuang Ho Hospital. This dataset included:

  • 30 patients with stroke.
  • 17 patients with occlusion but no stroke.
  • 3 patients with dementia.

A total of 616 images were used, with the dataset split into three parts: 10 patients for training, 10 for verification, and 10 for testing. Specifically, 205 images were used for training, 205 for verification, and 206 for testing. The ground-truth images were manually annotated by physicians, marking the lesion areas.

Key Insights and Ideas

  • Patch-based CNN for Lesion Detection: The decision to segment MRI images into smaller patches (7x7 pixels) was crucial. This approach effectively addresses the challenge of detecting very small lesions that might be lost or diluted in a whole-image analysis, allowing the CNN to learn localized features.
  • CPU-Optimized CNN: The authors intentionally designed their CNN to run on a CPU rather than a GPU. This choice prioritizes accessibility and lower computational cost, making the technology potentially more deployable in resource-constrained clinical settings, while still achieving high accuracy.
  • Importance of Preprocessing: The detailed preprocessing steps, particularly the removal of the cranial shell, are vital for focusing the model's attention solely on the brain tissue where lesions occur, thereby improving efficiency and accuracy.
  • Visualizing Model Output: The paper emphasizes the importance of visualizing the CNN's output directly on the original MRI image, overlaying the detected lesion areas. This direct visualization is key for clinical utility, allowing doctors to quickly assess the findings.

Key Results

The proposed CNN model achieved high accuracy in detecting cerebral small vessel lesions:

  • Training Accuracy: 98.57%
  • Verification Accuracy: 98.52%
  • Testing Accuracy: The paper reports specific accuracy figures for different lesion types in Figure 8, showing results like 98.83%, 96.23%, and 97.45% for different visual comparisons.
  • Comparison with MLP: The CNN model outperformed a Multilayer Perceptron (MLP) model in terms of lesion localization and detection of smaller lesions, as visually demonstrated in Figure 10.
  • Comparison with YOLO Models: While the paper presents tables for YOLO models (Tables 1-10), the direct comparison of F1-score, precision, and recall shows that the proposed method's F1-score for testing data (0.0454) is lower than YOLO models (e.g., YOLO3 F1-score of 0.1495). However, the precision rate for the proposed method (0.9590) is higher than YOLO models. This suggests a trade-off, where the proposed method might be more precise but less sensitive to detecting all possible lesions compared to YOLO.

Limitations and Future Work

  • Lesion Size: The authors acknowledge that detecting very small lesions remains a challenge, and the detected lesion boundaries might be slightly larger (about 2 pixels) than those marked by a physician.
  • Manual Preprocessing: The removal of the head bone, while effective, is currently a manual step that could be automated for greater efficiency.
  • Data Relevance for Stroke Grading: The current dataset and model may not fully capture all the data needed for comprehensive stroke association and grading.
  • Future Work: The authors propose several avenues for future research:
  • Automating the head bone removal process.
  • Improving the model's ability to detect smaller and more subtle lesions.
  • Incorporating more data to enhance the model's capacity for stroke association and grading.
  • Developing a 3D visualization of detected lesions to further assist clinicians in diagnosis.
  • Exploring deep neural network training for improved design stability and accuracy.

Summary updated

Abstract

Magnetic resonance imaging (MRI) offers the most detailed brain structure image available today; it can identify tiny lesions or cerebral cortical abnormalities. The primary purpose of the procedure is to confirm whether there is structural variation that causes epilepsy, such as hippocampal sclerotherapy, local cerebral cortical dysplasia, and cavernous hemangioma. Cerebrovascular disease, the second most common factor of death in the world, is also the fourth leading cause of death in Taiwan, with cerebrovascular disease having the highest rate of stroke. Among the most common are large vascular atherosclerotic lesions, small vascular lesions, and cardiac emboli. The purpose of this thesis is to establish a computer-aided diagnosis system based on small blood vessel lesions in MRI images, using the method of Convolutional Neural Network and deep learning to analyze brain vascular occlusion by analyzing brain MRI images. Blocks can help clinicians more quickly determine the probability and severity of stroke in patients. We analyzed MRI data from 50 patients, including 30 patients with stroke, 17 patients with occlusion but no stroke, and 3 patients with dementia. This system mainly helps doctors find out whether there are cerebral small vessel lesions in the brain MRI images, and to output the found results into labeled images. The marked contents include the position coordinates of the small blood vessel blockage, the block range, the area size, and if it may cause a stroke. Finally, all the MRI images of the patient are synthesized, showing a 3D display of the small blood vessels in the brain to assist the doctor in making a diagnosis or to provide accurate lesion location for the patient.
BibTeX
@inproceedings{hsieh2019cerebral,
  author = {Yi‐Zeng Hsieh and Yu Luo and Chen Pan and Mu Su and Chi Jen Chen and Kevin Li‐Chun Hsieh},
  title = {{Cerebral Small Vessel Disease Biomarkers Detection on MRI-Sensor-Based Image and Deep Learning}},
  booktitle = {Sensors},
  year = {2019},
  doi = {10.3390/s19112573},
}
Figure