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

Shangjie Yao, Yaowu Chen, Xiang Tian, Rongxin Jiang

TL;DR

This paper introduces DeepConv-DilatedNet, an improved deep learning algorithm for detecting and localizing pneumonia in chest X-rays. By enhancing the Faster R-CNN framework with a novel backbone and optimized post-processing, the method achieves state-of-the-art performance on benchmark datasets, offering crucial location information for clinical diagnosis.

What Problem Does This Paper Address?

Pneumonia is a significant global health threat, and timely diagnosis is critical. Chest X-rays (CXRs) are a primary diagnostic tool, but manual interpretation can be inefficient and prone to error due to the subtle and variable appearance of pneumonia lesions. Existing computer-aided diagnosis systems often struggle with accurately identifying and precisely locating these lesions, especially in low-contrast images. This paper aims to improve the accuracy and localization capabilities of automated pneumonia detection systems.

Methods

The core of this work is an improved detection framework built upon the Faster R-CNN architecture, a well-established two-stage object detector. The key innovations lie in:

  • DeepConv-DilatedNet Backbone: This novel backbone replaces the standard convolutional layers with a dilated bottleneck residual neural network. Dilated convolutions allow for a larger receptive field without increasing computational cost or losing spatial resolution, which is crucial for capturing subtle features of pneumonia. The network also incorporates deconvolution (instead of upsampling) to better preserve feature information when restoring feature maps to their original size.
  • Feature Pyramid Network (FPN) Integration: FPN is integrated to combine low-level and high-level features, enabling the detection of objects at multiple scales.
  • Soft-NMS for Proposal Filtering: To address the issue of overlapping bounding boxes generated by the Region Proposal Network (RPN), Soft-Non-Maximum Suppression (Soft-NMS) is employed. Unlike traditional NMS which hard-deletes boxes, Soft-NMS reduces the scores of overlapping boxes, allowing for better detection of closely spaced or overlapping lesions.
  • K-Means++ for Anchor Box Optimization: The authors use the K-Means++ algorithm to determine optimal anchor box sizes and aspect ratios tailored to the specific characteristics of pneumonia lesions in CXRs, improving localization accuracy.
  • CLAHE Preprocessing: To enhance the visibility of pneumonia lesions, which often have low contrast, the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm is applied as a preprocessing step.

Datasets

The study utilizes two publicly available datasets for training and evaluation:

  • Radiological Society of North America (RSNA) Pneumonia Detection Dataset: This dataset contains chest X-ray images with radiologist annotations for pneumonia. The authors preprocessed this dataset by filtering out irrelevant images and augmenting the remaining ones.
  • ChestX-ray14 Dataset: This dataset, also publicly available, was used for further evaluation. It contains a larger number of images and pathology classes, including pneumonia.

Key Insights and Ideas

  • Dilated Convolutions are Key for Subtle Features: The paper strongly suggests that dilated convolutions are highly effective in capturing the nuanced visual cues of pneumonia in CXRs, which are often obscured by low contrast and subtle texture changes.
  • Deconvolution Preserves More Information: Replacing standard upsampling with deconvolution in the feature pyramid construction helps retain more valuable spatial and semantic information from high-level feature maps, leading to better localization.
  • Soft-NMS is Crucial for Overlapping Lesions: The effectiveness of Soft-NMS in handling overlapping bounding boxes is highlighted, demonstrating its importance for accurate detection when multiple pneumonia lesions are present or close together.
  • Dataset-Specific Anchor Boxes Matter: Optimizing anchor boxes using K-Means++ significantly improves the model's ability to accurately frame the pneumonia targets, rather than relying on generic or empirically chosen sizes.

Key Results

The proposed DeepConv-DilatedNet-based Faster R-CNN model achieved state-of-the-art performance on the benchmark datasets:

  • RSNA Dataset: Achieved 39.23% Mean Average Precision (mAP).
  • ChestX-ray14 Dataset: Achieved 38.02% mAP.
  • The model consistently outperformed other benchmark networks (DetNet59, ResNet50, ResNet101, VGG16) across various Intersection over Union (IoU) thresholds.
  • Qualitative results (Figure 11) show that the model's bounding boxes are more accurate and better fitted to the pneumonia lesions compared to other models.
  • ROC curves indicate a higher AUC for the proposed model, signifying better classification performance.

Limitations and Future Work

  • Dataset Size and Quality: While the RSNA dataset is substantial, the authors note that initial filtering was necessary due to the presence of many "useless images." The ChestX-ray14 dataset, while larger, had a very small number of pneumonia images with bounding box annotations (120), which might limit the model's generalization on this specific subset.
  • Computational Cost: Although efforts were made to optimize the network (e.g., using dilated convolutions), deep learning models for medical imaging can still be computationally intensive for training and inference.
  • Generalizability to Other Diseases: The current work focuses specifically on pneumonia. Future work could explore extending this architecture to detect other lung pathologies or diseases in different medical imaging modalities.
  • Clinical Validation: While the model shows promising results, further validation in real-world clinical settings with diverse patient populations and imaging equipment would be essential.

Summary updated

Abstract

Pneumonia remains a threat to human health; the coronavirus disease 2019 (COVID-19) that began at the end of 2019 had a major impact on the world. It is still raging in many countries and has caused great losses to people’s lives and property. In this paper, we present a method based on DeepConv-DilatedNet of identifying and localizing pneumonia in chest X-ray (CXR) images. Two-stage detector Faster R-CNN is adopted as the structure of a network. Feature Pyramid Network (FPN) is integrated into the residual neural network of a dilated bottleneck so that the deep features are expanded to preserve the deep feature and position information of the object. In the case of DeepConv-DilatedNet, the deconvolution network is used to restore high-level feature maps into its original size, and the target information is further retained. On the other hand, DeepConv-DilatedNet uses a popular fully convolution architecture with computation shared on the entire image. Then, Soft-NMS is used to screen boxes and ensure sample quality. Also, K-Means++ is used to generate anchor boxes to improve the localization accuracy. The algorithm obtained 39.23% Mean Average Precision (mAP) on the X-ray image dataset from the Radiological Society of North America (RSNA) and got 38.02% Mean Average Precision (mAP) on the ChestX-ray14 dataset, surpassing other detection algorithms. So, in this paper, an improved algorithm that can provide doctors with location information of pneumonia lesions is proposed.
BibTeX
@inproceedings{yao2021pneumonia,
  author = {Shangjie Yao and Yaowu Chen and Xiang Tian and Rongxin Jiang},
  title = {{Pneumonia Detection Using an Improved Algorithm Based on Faster R-CNN}},
  booktitle = {Comput. Math. Methods Medicine},
  year = {2021},
  url = {https://openreview.net/forum?id=MMRbHGPvBcl},
}
Figure