Pneumonia Detection Using an Improved Algorithm Based on Faster R-CNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pneumonia, including COVID-19, poses a significant global health threat.
• Accurate and timely diagnosis of pneumonia from chest X-ray (CXR) images is crucial.
• Existing detection methods require improvement in localization accuracy.

Purpose of the Study:

• To develop and evaluate a novel deep learning method for identifying and localizing pneumonia in CXR images.
• To improve the accuracy of pneumonia detection and localization compared to existing algorithms.
• To provide clinicians with precise location information of pneumonia lesions.

Main Methods:

• A two-stage detector, Faster R-CNN, forms the network's core structure.
• Feature Pyramid Network (FPN) integration enhances deep feature preservation and positional information.
• DeepConv-DilatedNet utilizes deconvolution for feature map restoration and a fully convolutional architecture for shared computation.
• Soft-NMS and K-Means++ are employed for sample quality screening and anchor box optimization, respectively.

Main Results:

• The DeepConv-DilatedNet algorithm achieved 39.23% Mean Average Precision (mAP) on the RSNA CXR dataset.
• The method obtained 38.02% mAP on the ChestX-ray14 dataset.
• Performance surpassed other existing pneumonia detection algorithms.

Conclusions:

• The proposed DeepConv-DilatedNet offers an effective solution for pneumonia identification and localization in CXR images.
• The algorithm demonstrates superior performance, outperforming other detection methods.
• This advancement can assist doctors by providing critical location data for pneumonia lesions.