Detection of acute thoracic aortic dissection based on plain chest radiography and a residual neural network (Resnet)

Dong Keon Lee, Jin Hyuk Kim, Jaehoon Oh, Tae Hyun Kim, Myeong Seong Yoon, Dong Jin Im, Jae Ho Chung, Hayoung ByunScientific 2022 · DOI

TL;DR

This paper presents a deep learning model, specifically a ResNet18, that can detect acute thoracic aortic dissection from standard chest X-rays with 90.20% accuracy. This approach offers a faster and more accessible screening tool compared to current methods, potentially improving patient outcomes for this life-threatening condition.

What Problem Does This Paper Address?

Acute thoracic aortic dissection is a critical medical emergency with a high mortality rate if not diagnosed and treated promptly. While contrast-enhanced CT scans are the gold standard for diagnosis, they are not always immediately available or feasible due to factors like kidney function and cost. Chest X-rays are commonly used as an initial screening tool due to their accessibility and low radiation dose, but their diagnostic accuracy for aortic dissection is relatively low (around 70% sensitivity). This gap highlights the need for a more accurate and efficient screening method that can leverage readily available chest X-ray data.

Methods

The researchers employed a Residual Neural Network (ResNet18) for image classification. ResNet architectures are known for their ability to handle deep networks by using "skip connections" to mitigate the vanishing gradient problem. The study involved several key steps:

  • Data Pre-processing and Augmentation: Chest X-ray images were standardized by removing personal information, cropping, resizing to 448x448 pixels, and applying histogram equalization to improve contrast. To increase the robustness and size of the training dataset, augmentation techniques like random flipping and rotation were used.
  • Model Architecture: ResNet18 was chosen for its balance of performance and parameter efficiency compared to other models like DenseNet and EfficientNet.
  • Handling Data Imbalance: The dataset had significantly more negative (normal) images than positive (dissection) images. This imbalance was addressed using a weighted random sampling method and a weighted cross-entropy loss function, ensuring that the model paid adequate attention to the minority class.
  • Ensemble Voting: To further enhance robustness and prevent overfitting, a fivefold cross-validation approach was used, and the predictions from five models were combined using a soft voting system.
  • Visual Verification: Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to visualize which regions of the X-ray images the ResNet18 model focused on for its predictions, providing insights into its decision-making process.

Datasets

The study utilized a multi-center dataset comprising 3,331 chest X-ray images from three tertiary academic hospitals in South Korea.

  • Training and Validation Data: Images from Seoul National University Bundang Hospital (Hospital A) and Hanyang University Hospital (Hospital B) were used for training and internal validation. Hospital A contributed 1,972 images (507 positive), and Hospital B contributed 1,155 images (155 positive). The training set comprised 91% of the data, while 9% was allocated to the testing set from Hospital C.
  • Testing Data: Images from Yonsei University Hospital (Hospital C) were used exclusively for testing the model's performance on unseen data. Hospital C provided 204 images (54 positive).
  • Data Split: The dataset consisted of 716 positive images (acute thoracic aortic dissection) and 2,615 negative images (normal).

Key Insights and Ideas

  • Leveraging Standard X-rays: The core insight is that even standard, low-cost chest X-rays, when analyzed by a sophisticated deep learning model, can achieve diagnostic performance comparable to more advanced imaging modalities for acute thoracic aortic dissection.
  • ResNet18's Efficiency: The choice of ResNet18 proved effective, offering a good balance between model complexity and diagnostic accuracy, outperforming other tested deep learning architectures.
  • Importance of Data Handling: Effectively addressing data imbalance and using ensemble methods were crucial for achieving robust and reliable performance, preventing the model from being biased towards the majority class.
  • Visualizing Model Attention: Grad-CAM revealed that the model correctly focused on the aortic region and related features (like mediastinal widening) for positive cases, providing a degree of interpretability.

Key Results

The ResNet18 model achieved the following performance metrics on the testing dataset:

  • Diagnostic Accuracy: 90.20%
  • Precision: 75.00%
  • Recall (Sensitivity): 94.44%
  • F1-score: 83.61%

The Area Under the Receiver Operating Characteristic Curve (AUC) for the trained model on the testing dataset was 0.955, indicating excellent discrimination ability. The study also noted that the model's performance was comparable to or better than other diagnostic tools like TEE and CT in terms of accuracy, and better than TTE.

Limitations and Future Work

The authors acknowledge several limitations:

  • Age and Sex Matching: The patient demographics (age and sex) were not perfectly matched between dissection and non-dissection groups, which could influence results as aortic dissection prevalence varies with these factors.
  • Image Variability: Input images were of assorted sizes and pre-processed, but variations in imaging environments might still affect performance.
  • Imbalanced Test Dataset: While the training data was balanced, the test dataset might still exhibit some imbalance, potentially affecting reliability.
  • Binary Classification: The model only performed binary classification (normal vs. dissection) and did not differentiate between types of aortic dissection (e.g., Stanford Type A or B) or other chest pain causes.
  • Replacement for CT: The authors emphasize that while promising, these results are not sufficient to conclude that chest X-rays can completely replace CT scans for suspected aortic dissection.

Future work suggested includes:

  • Aorta Segmentation: Improving accuracy by incorporating aorta segmentation into the model.
  • Clinical Applicability: Developing research directly applicable to clinical practice.
  • Multi-class Classification: Expanding the model to classify different types of aortic dissection and other chest pathologies.

Summary updated

Abstract

Acute thoracic aortic dissection is a life-threatening disease, in which blood leaking from the damaged inner layer of the aorta causes dissection between the intimal and adventitial layers. The diagnosis of this disease is challenging. Chest x-rays are usually performed for initial screening or diagnosis, but the diagnostic accuracy of this method is not high. Recently, deep learning has been successfully applied in multiple medical image analysis tasks. In this paper, we attempt to increase the accuracy of diagnosis of acute thoracic aortic dissection based on chest x-rays by applying deep learning techniques. In aggregate, 3,331 images, comprising 716 positive images and 2615 negative images, were collected from 3,331 patients. Residual neural network 18 was used to detect acute thoracic aortic dissection. The diagnostic accuracy of the ResNet18 was observed to be 90.20% with a precision of 75.00%, recall of 94.44%, and F1-score of 83.61%. Further research is required to improve diagnostic accuracy based on aorta segmentation.
BibTeX
@inproceedings{lee2022detection,
  author = {Dong Keon Lee and Jin Hyuk Kim and Jaehoon Oh and Tae Hyun Kim and Myeong Seong Yoon and Dong Jin Im and Jae Ho Chung and Hayoung Byun},
  title = {{Detection of acute thoracic aortic dissection based on plain chest radiography and a residual neural network (Resnet)}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-26486-3},
}
Figure