COVID-19 Detection Using CT Image Based On YOLOv5 Network

Ruyi Qu, Yi Yang, Yuwei Wang

TL;DR

This paper tackles the critical need for rapid and accurate COVID-19 detection using CT scans by applying the YOLOv5 object detection model. The authors demonstrate that YOLOv5 achieves competitive performance, outperforming established models like Faster RCNN and EfficientDet in detecting COVID-19 abnormalities on chest radiographs.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the urgent need for efficient and timely medical diagnostics. Traditional methods like PCR tests can be time-consuming, and diagnosing COVID-19 from chest radiographs can be challenging due to its similarity to other forms of pneumonia. Computer-aided diagnosis (CAD) systems are crucial for improving efficiency and accuracy, especially in image-based diagnostics like CT scans. This paper focuses on developing a CAD system for detecting and classifying COVID-19 abnormalities in CT images, aiming to assist radiologists in making quicker and more confident diagnoses.

Methods

The core of this work is the application of the YOLOv5 object detection network. YOLOv5 is a single-stage object detection model known for its speed and efficiency. The authors describe its architecture, which includes:

  • Backbone (CSPDarknet): This component is responsible for extracting features from the input CT images. It utilizes CSPNet technology to improve gradient flow and reduce model complexity, leading to better accuracy and efficiency.
  • Neck (PANet): This part enhances feature propagation by integrating bottom-up and top-down pathways, improving the network's ability to capture multi-scale features.
  • Head (Yolo Layer): This is where the final detection and classification occur, predicting bounding boxes and class probabilities for COVID-19 abnormalities.

The paper also briefly touches upon the broader landscape of object detection, categorizing methods into two main streams: two-stage (e.g., Faster RCNN) and one-stage (e.g., YOLO series).

Datasets

The study utilizes a Kaggle-provided dataset for training and evaluation. This dataset consists of 6364 chest scans in DICOM format, with labels provided by experienced radiologists. The labels categorize scans into three types: negative for pneumonia, typical COVID-19 appearance, and indeterminate/atypical COVID-19 appearance. The paper also mentions a train_study_level file containing study identifiers and their corresponding labels.

Key Insights and Ideas

  • YOLOv5's Efficacy in Medical Imaging: The paper demonstrates that a state-of-the-art object detection model, typically used in general computer vision tasks, can be effectively adapted for medical image analysis, specifically for COVID-19 detection.
  • Single-Stage vs. Two-Stage Performance: By comparing YOLOv5 (single-stage) with Faster RCNN (two-stage), the study implicitly supports the growing trend of single-stage detectors offering a compelling balance of speed and accuracy, even in specialized domains.
  • Importance of mAP for Object Detection: The paper emphasizes mean Average Precision (mAP) as the primary metric for evaluating object detection performance, highlighting its role in assessing both the accuracy of bounding box predictions and the correctness of classifications.

Key Results

The YOLOv5s model achieved a mean Average Precision (mAP) of 0.623 at an IoU threshold of 0.5 ([email protected]). This performance was superior to:

Specifically, YOLOv5s outperformed Faster RCNN by 0.157 and EfficientDet by 0.101 in terms of [email protected]. The authors also note that the chest is the most frequently examined body part in their dataset.

Limitations and Future Work

  • Model Variants: The paper focuses on the YOLOv5s variant, which is the smallest and fastest in the YOLOv5 family. The authors mention that other YOLOv5 variants exist (e.g., YOLOv5m, YOLOv5l, YOLOv5x) which are deeper and wider, suggesting potential for further performance improvements by exploring these.
  • Dataset Size and Diversity: While the dataset size is substantial (6364 scans), the paper doesn't extensively discuss potential biases or limitations related to the diversity of patient populations, scanner types, or image acquisition protocols.
  • Further Adjustments: The conclusion section mentions the proposal of "further adjustment in the future," but specific details on these adjustments are not elaborated upon in the provided text.

Summary updated

Abstract

The COVID-19 pandemic has broken down the global medical order tremendously, we urgently need an efficient treatment. Computer aided diagnosis (CAD) increases diagnosis efficiency, helping doctors providing a quick and confident diagnosis, it has played an important role in the treatment of COVID-19. In our task, we solve the problem about abnormality detection and classification. The dataset provided by Kaggle platform and we choose YOLOv5 as our model.We introduce some methods on objective detection in the related work section, the objection detection can be divided into two streams: one-stage and two stage. The representational model are Faster RCNN and YOLO series. Then in the section III we describe YOLOv5 model in the detail. Compared Experiment and results are shown in section IV. We choose mean average precision (mAP) as our experiments’ metrics, and the higher (mean )mAP is, the better result the model will gain. [email protected] of our YOLOv5s is 0.623 which is 0.157 and 0.101 higher than Faster RCNN and EfficientDet respectively.
BibTeX
@inproceedings{qu2021covid,
  author = {Ruyi Qu and Yi Yang and Yuwei Wang},
  title = {{COVID-19 Detection Using CT Image Based On YOLOv5 Network}},
  booktitle = {2021},
  year = {2021},
  doi = {10.1109/iaecst54258.2021.9695714},
}
Figure