Proposing a novel deep network for detecting COVID-19 based on chest images

Maryam Dialameh, Ali Hamzeh, Hossein Rahmani, Amir Reza Radmard, Safoura DialamehScientific 2022 · DOI

TL;DR

Researchers have introduced CT-COV19, a massive new dataset of over 13,000 CT images, alongside Deep-CT-Net, a neural network utilizing a pixel-wise "pyramid attention" layer to significantly boost sensitivity in COVID-19 detection. The study also demonstrates a transfer-learning extension for X-rays (Deep-CXR-Net) that achieves an impressive 0.98 AUC, providing a robust framework for rapid, automated screening.

What Problem Does This Paper Address?

While RT-PCR remains the gold standard for COVID-19 diagnosis, it suffers from high false-negative rates and logistical bottlenecks. Chest imaging (CT and X-ray) offers a faster alternative, but manual review is exhausting for radiologists during pandemic waves.

Existing AI models for chest imaging have faced two major hurdles:

  1. Data Scarcity and Bias: Most models are trained on small datasets often biased toward specific age groups or regions, leading to poor generalization.
  2. Low Sensitivity: Many models struggle to detect subtle infection patterns during the early incubation period, leading to dangerous false negatives.

The authors aim to solve these by providing a large-scale, diverse dataset and a model architecture specifically designed to "pay attention" to subtle, pixel-level indicators of infection.

Methods

The researchers proposed two primary architectures:

  1. Deep-CT-Net (for CT scans): This model uses DenseNet-121 as a backbone for feature extraction. The "secret sauce" is the addition of a Pyramid Attention Layer on top of high-level features. Unlike standard global pooling, this layer provides pixel-wise attention, allowing the model to focus on specific infectious regions within the lungs that might be overlooked by the human eye or simpler models.
  2. Deep-CXR-Net (for X-rays): Recognizing that COVID-19 X-ray data is scarce, the authors used a transfer learning strategy. They built a three-part ensemble: two branches pre-trained on massive non-COVID datasets (CheXpert and Kaggle-Pneumonia) and a third branch trained on COVID-19 samples. By concatenating these features, the model leverages knowledge of general lung diseases to better identify COVID-specific anomalies.

Notably, the workflow avoids heavy pre-processing steps like manual lung segmentation, which speeds up the inference time significantly—a critical factor for emergency room use.

Datasets

The paper introduces a significant new resource:

  • CT-COV19 (New): A publicly available dataset of 13,505 CT images from over 1,000 patients. It is notably diverse, including patients aged 19 to 73 from four different geographic regions with varying climates. It includes 8,500 COVID-19 images, 754 "other pneumonia" images, and 4,251 normal images.

For validation and generalization testing, they used:

  • COVID-CT: A common benchmark dataset.
  • ieee8023: A dataset used to evaluate the X-ray (CXR) extension.

Key Insights and Ideas

  • Sensitivity over Precision: In a pandemic, a false negative (missing a sick person) is often more dangerous than a false positive (extra testing for a healthy person). The authors deliberately optimized for sensitivity, achieving 0.858 on CT scans compared to the 0.81 of previous state-of-the-art models.
  • The Power of Attention: The Pyramid Attention layer proved crucial. By visualizing the model's focus via Class Activation Mapping (CAM), the authors showed that the network accurately "looks" at the specific inflammatory regions of the lungs to make its decision.
  • Generalization is Key: A major "aha" moment was testing Deep-CT-Net on the external COVID-CT dataset without any fine-tuning. It achieved an AUC of 0.92, proving that the diversity of their new CT-COV19 dataset helps the model learn universal features rather than just memorizing a specific hospital's data.

Key Results

  • CT Performance: On the new CT-COV19 dataset, the model achieved an AUC of 0.886 and a sensitivity of 0.858, outperforming competitors like COVNET and Hybrid-3D.
  • X-Ray Performance: The Deep-CXR-Net extension reached a staggering AUC of 0.9839 and a sensitivity of 0.9824 on the ieee8023 benchmark.
  • Efficiency: The model is lightweight enough to run on a standard computer with an ordinary consumer-grade GPU (like an Nvidia card), making it viable for hospitals without massive server clusters.

Limitations and Future Work

The authors acknowledge a few key areas for improvement:

  • Precision Trade-off: While sensitivity is high, the precision is slightly lower than some competitors. This means the model may produce more false positives, requiring a doctor's secondary review.
  • Multi-class Classification: Currently, the model is primarily a binary classifier (COVID vs. Normal). It struggles to specifically quantify or distinguish between different types of non-COVID bacterial pneumonias.
  • Future Decoder Integration: The authors plan to develop a "deep neural decoder" to not just classify the image, but automatically segment (outline) the infected parts of the lung, which could help doctors track disease progression over time.

Summary updated

Abstract

The rapid outbreak of coronavirus threatens humans' life all around the world. Due to the insufficient diagnostic infrastructures, developing an accurate, efficient, inexpensive, and quick diagnostic tool is of great importance. To date, researchers have proposed several detection models based on chest imaging analysis, primarily based on deep neural networks; however, none of which could achieve a reliable and highly sensitive performance yet. Therefore, the nature of this study is primary epidemiological research that aims to overcome the limitations mentioned above by proposing a large-scale publicly available dataset of chest computed tomography scan (CT-scan) images consisting of more than 13k samples. Secondly, we propose a more sensitive deep neural networks model for CT-scan images of the lungs, providing a pixel-wise attention layer on top of the high-level features extracted from the network. Moreover, the proposed model is extended through a transfer learning approach for being applicable in the case of chest X-Ray (CXR) images. The proposed model and its extension have been trained and evaluated through several experiments. The inclusion criteria were patients with suspected PE and positive real-time reverse-transcription polymerase chain reaction (RT-PCR) for SARS-CoV-2. The exclusion criteria were negative or inconclusive RT-PCR and other chest CT indications. Our model achieves an AUC score of 0.886, significantly better than its closest competitor, whose AUC is 0.843. Moreover, the obtained results on another commonly-used benchmark show an AUC of 0.899, outperforming related models. Additionally, the sensitivity of our model is 0.858, while that of its closest competitor is 0.81, explaining the efficiency of pixel-wise attention strategy in detecting coronavirus. Our promising results and the efficiency of the models imply that the proposed models can be considered reliable tools for assisting doctors in detecting coronavirus.
BibTeX
@inproceedings{dialameh2022proposing,
  author = {Maryam Dialameh and Ali Hamzeh and Hossein Rahmani and Amir Reza Radmard and Safoura Dialameh},
  title = {{Proposing a novel deep network for detecting COVID-19 based on chest images}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-06802-7},
}
Figure