Attentional decoder networks for chest X-ray image recognition on high-resolution features.

Area of Science:

• Medical Imaging Analysis
• Deep Learning for Radiology
• Computer-Aided Diagnosis

Background:

• Small lesions in chest X-rays are often missed due to down-sampling in encoder-only networks.
• Existing U-Net architectures face challenges in resource-intensive up-sampling and effective high-resolution feature pooling for classification.
• Accurate detection of subtle pathological findings is crucial for early disease diagnosis.

Purpose of the Study:

• To introduce an encoder-decoder network specifically designed for recognizing small-size lesions in chest X-ray images.
• To overcome the limitations of existing methods in handling high-resolution feature maps for lesion classification.
• To improve the efficiency and accuracy of deep learning models in medical image analysis.

Main Methods:

• Proposed an encoder-decoder network incorporating a lightweight attentional decoder for efficient feature up-sampling.
• Utilized harmonic magnitude transform for pooling high-resolution feature maps in the frequency domain, preserving translation invariance.
• Employed an efficient embedding strategy to reduce parameters in the pooling layer.

Main Results:

• The proposed network achieved state-of-the-art classification performance on three public chest X-ray datasets (NIH, CheXpert, MIMIC-CXR).
• Demonstrated effective recognition of small-size lesions, addressing a key challenge in chest X-ray analysis.
• The model's ability to maintain pathological locality while incorporating global context was validated.

Conclusions:

• The developed encoder-decoder network efficiently recognizes small lesions in chest X-rays.
• The combination of an attentional decoder and harmonic magnitude transform offers a promising solution for lesion detection.
• The open-source implementation facilitates further research and development in medical image analysis.