A Review of Automated Diagnosis of COVID-19 Based on Scanning Images

Delong Chen, Shunhui Ji, Fan Liu, Zewen Li, Xinyu Zhou

TL;DR

This paper provides a comprehensive review of 70 automated COVID-19 diagnosis models developed using machine learning and deep learning on medical imaging data. It analyzes these models across preprocessing, feature extraction, classification, and evaluation stages, highlighting current trends and identifying domain adaptation and interpretability as crucial future research directions.

What Problem Does This Paper Address?

The COVID-19 pandemic's rapid spread created an urgent need for efficient and accurate diagnostic tools. Traditional RT-PCR tests are time-consuming and can yield false negatives. Medical imaging, specifically X-rays and CT scans, emerged as a promising alternative for faster diagnosis. However, the sheer volume of images and the burden on radiologists necessitated the development of automated diagnostic systems. This paper addresses the challenge of synthesizing the rapidly growing body of research in this area, providing a structured overview of existing automated diagnosis models, their methodologies, and their limitations.

Methods

The authors systematically reviewed 70 COVID-19 diagnosis models proposed between February 14 and July 21, 2020. They analyzed these models by breaking down their pipelines into four key stages:

  • Preprocessing: This stage involves techniques to prepare the medical images for analysis. Common methods include data augmentation (rotating, flipping, scaling, etc.) to increase dataset size and prevent overfitting, histogram equalization to improve image contrast, and lung segmentation to isolate the region of interest and reduce interference from surrounding tissues. Advanced augmentation techniques like Generative Adversarial Networks (GANs) were also explored.
  • Feature Extraction: This is where discriminative lesion patterns are identified. The vast majority of models utilize Convolutional Neural Networks (CNNs), with popular architectures like ResNet, GoogLeNet, and DenseNet being frequently employed. Some researchers also explored automatic network design and model ensemble methods to enhance feature extraction. Novel approaches included using local patches, multi-scale learning, and 3D CNNs to capture spatial information more effectively.
  • Classification: Once features are extracted, a classifier predicts the presence of COVID-19. Most models use CNNs with a softmax classifier. Some papers propose hybrid approaches combining softmax with decision trees or other ensemble methods. Non-end-to-end models employing Support Vector Machines (SVMs) were also noted. The paper discusses both binary (COVID-19 vs. normal) and multi-class classification strategies.
  • Evaluation: Models are assessed using metrics like accuracy and Area Under the Curve (AUC). The paper also notes the importance of considering dataset size and class imbalance in evaluation.

Datasets

The paper analyzes models trained on both X-ray and CT scan datasets. Specific dataset sizes are detailed in Table II, which lists the number of COVID-19 and total samples in training and testing sets for various studies. The authors highlight that X-ray training sets are generally larger than CT scan training sets, which might contribute to the surprisingly better performance of X-ray-based models in some cases. The paper also points out that many datasets are private, making direct comparison and reproduction of results challenging.

Key Insights and Ideas

  • The rapid evolution of AI for COVID-19 diagnosis: The sheer number of models developed in a short period (February-July 2020) underscores the urgency and intense research effort in this area.
  • X-ray models outperform CT models surprisingly: Despite CT scans theoretically offering more detailed information, X-ray-based models often achieve higher accuracy. This is attributed to the larger availability of X-ray training data.
  • The challenge of imbalanced datasets: Many studies suffer from highly imbalanced test sets, where COVID-19 cases are significantly outnumbered by normal cases, potentially inflating performance metrics.
  • Transfer learning is a double-edged sword: While crucial for overcoming limited data, pre-training on general datasets like ImageNet may not transfer optimal features for medical imaging. Fine-tuning on domain-specific datasets (e.g., lung cancer) or more targeted transfer learning approaches are explored.
  • Interpretability is key for clinical adoption: While models can achieve high accuracy, understanding why a model makes a certain prediction is vital for radiologists. Techniques like Grad-CAM are used, but more advanced methods are needed.
  • Detection-based models offer inherent interpretability: Directly outputting bounding boxes or masks for lesions provides a more intuitive understanding of the diagnosis compared to classification alone.

Key Results

The paper summarizes evaluation results in Table II. On average, X-ray scanning-based models achieved an accuracy of 94.76% and an AUC of 96.94%, while CT scanning-based models achieved an average accuracy of 90.13% and an AUC of 94.76%. The paper notes that while many models claim high performance, the size and balance of their test sets are often a concern.

Limitations and Future Work

The authors identify several limitations and suggest future research directions:

  • Domain Adaptation: Current transfer learning approaches often rely on pre-training on general datasets. Developing deep domain adaptation techniques to better bridge the gap between source and target domains (e.g., general images to medical images) is crucial for leveraging limited COVID-19 data effectively.
  • Interpretability Promotion: While methods like Grad-CAM offer some insight, there's a need for more robust interpretability techniques. Detection-based models, which directly localize lesions, are highlighted as a promising avenue for improving interpretability and clinical utility.
  • Benchmark Datasets: The lack of standardized, publicly available benchmark datasets makes it difficult to compare models rigorously and reproduce results. The authors emphasize the need for such benchmarks.
  • Generalizability: The models reviewed were developed during the early stages of the pandemic. Their performance on new variants or different patient populations may vary, necessitating further research into robust and generalizable models.

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 Shunhui Ji and Fan Liu and Zewen Li and Xinyu Zhou},
  title = {{A Review of Automated Diagnosis of COVID-19 Based on Scanning Images}},
  booktitle = {Europe},
  year = {2020},
}
Figure