A Review of Automatically Diagnosing COVID-19 based on Scanning Image

Delong Chen, Fan Liu, Zewen Li

TL;DR

This paper reviews 70 automated COVID-19 diagnosis models developed using machine learning and deep learning on X-ray and CT scans between February and July 2020. It analyzes these models across preprocessing, feature extraction, classification, and evaluation stages, highlighting the rapid advancements and identifying key challenges and future research directions in domain adaptation and interpretability.

What Problem Does This Paper Address?

The COVID-19 pandemic presented a critical need for rapid and accurate diagnosis. Traditional RT-PCR tests are time-consuming and can yield false negatives. This led to a surge in research using medical imaging (X-ray and CT scans) for diagnosis. However, the sheer volume of research and the rapid development of new models created a need for a comprehensive overview to understand the landscape, identify common approaches, and pinpoint areas for future improvement. This paper aims to fill that gap by systematically reviewing and analyzing the existing automated diagnostic models.

Methods

The authors provide a structured review of 70 COVID-19 diagnostic models, categorizing them by their pipeline stages:

  • Preprocessing: This includes data augmentation (rotation, flipping, scaling, etc.), advanced techniques like Generative Adversarial Networks (GANs) for generating virtual samples, histogram equalization for image enhancement, and lung segmentation (using U-Net variants, Dense-Net, etc.) to isolate the region of interest.
  • Feature Extraction: Most models leverage pre-trained Convolutional Neural Networks (CNNs) like ResNet, GoogLeNet, and DenseNet. Some researchers explored automatic network design, model ensembles, and local feature extraction methods to capture discriminative patterns. Multi-scale learning and 3D CNNs were also employed to address varying lesion sizes and the 3D nature of CT scans.
  • Classification: The primary approach is CNNs followed by softmax classifiers. Variations include combining softmax with decision trees or Adaboost. Non-end-to-end models using Support Vector Machines (SVMs) were also noted. The paper discusses both binary (COVID-19 vs. normal) and multi-class classification strategies.
  • Evaluation: Standard metrics like accuracy and Area Under the Curve (AUC) are used. The paper also notes the use of heatmaps (e.g., Grad-CAM) for interpretability and bounding boxes/masks in detection-based models.

Datasets

The paper analyzes models trained on various datasets, primarily X-ray and CT scans. Specific datasets are not explicitly named as benchmarks, but the authors highlight the size differences: the average training set size for X-ray models is 4185, while for CT models, it's only 1417. They also point out the issue of imbalanced test sets, with an average COVID-19 to total sample ratio of 0.274:1. The authors suggest the need for a proper benchmark test set for future research.

Key Insights and Ideas

  • X-ray vs. CT Performance Surprise: Despite CT scans theoretically containing more information, X-ray-based models surprisingly achieved better average performance. The authors attribute this to the larger training datasets available for X-ray images.
  • The Challenge of Fine-Grained Classification: Diagnosing COVID-19 is a fine-grained task, as different lung diseases can share similar visual patterns, leading to misclassification. This highlights the need for more sophisticated classification strategies.
  • Transfer Learning Dominance, But with Caveats: Transfer learning, particularly fine-tuning models pre-trained on ImageNet, is a prevalent strategy to overcome limited COVID-19 data. However, the authors note that ImageNet features might not be optimal for medical images, suggesting pre-training on related medical datasets (like lung cancer) could be more effective.
  • Interpretability is Crucial: While deep learning models offer high accuracy, their "black-box" nature is a limitation. Techniques like Grad-CAM are used to provide heatmaps, but the paper emphasizes the need for further advancements in interpretability to truly assist radiologists.
  • Domain Adaptation as a Promising Frontier: The authors identify domain adaptation as a key future direction, especially for addressing the scarcity of high-quality, large-scale labeled COVID-19 data.

Key Results

The paper summarizes evaluation results, noting:

  • Average accuracy for X-ray based models: 94.76%
  • Average AUC for X-ray based models: 96.94%
  • Average accuracy for CT scanning based models: 90.13%
  • Average AUC for CT scanning based models: 94.76%

These figures represent the performance of models reviewed in the paper, with the caveat that many models were evaluated on private datasets and with imbalanced test sets.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Limited Dataset Size and Imbalance: Many models are trained and tested on small, imbalanced datasets, raising concerns about generalizability and overfitting. The lack of standardized benchmark datasets hinders direct comparison.
  • Domain Adaptation: The paper strongly advocates for research into deep domain adaptation techniques to better leverage knowledge from related domains and improve performance on limited COVID-19 data.
  • Interpretability Promotion: Enhancing model interpretability is crucial for clinical adoption. The authors suggest that detection-based models offer inherent advantages and that further research in this area is vital.
  • Transfer Learning Optimization: While transfer learning is widely used, optimizing the pre-training and fine-tuning strategies, potentially by using more relevant source domains, is an area for improvement.

Summary updated

Abstract

The pandemic of COVID-19 has caused millions of infections, which has led to a great loss all over the world, socially and economically. Due to the false-negative rate and the time-consuming of the conventional Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests, diagnosing based on X-ray images and Computed Tomography (CT) images has been widely adopted. Therefore, researchers of the computer vision area have developed many automatic diagnosing models based on machine learning or deep learning to assist the radiologists and improve the diagnosing accuracy. In this paper, we present a review of these recently emerging automatic diagnosing models. 70 models proposed from February 14, 2020, to July 21, 2020, are involved. We analyzed the models from the perspective of preprocessing, feature extraction, classification, and evaluation. Based on the limitation of existing models, we pointed out that domain adaption in transfer learning and interpretability promotion would be the possible future directions.
BibTeX
@inproceedings{chen2020review,
  author = {Delong Chen and Fan Liu and Zewen Li},
  title = {{A Review of Automatically Diagnosing COVID-19 based on Scanning Image}},
  booktitle = {Europe},
  year = {2020},
}
Figure