Segmentation-Assisted Fusion-Based Classification for Automated CXR Image Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate classification of chest X-rays (CXRs) is vital for diagnosing lung diseases.
• Existing deep learning models struggle to differentiate non-lung features in CXR images.

Purpose of the Study:

• To propose a novel segmentation-assisted fusion-based classification method for CXR images.
• To enhance the accuracy and efficiency of lung disease diagnosis using deep learning.

Main Methods:

• A lightweight segmentation model, Partial Convolutional Segmentation Network (PCSNet), was developed using an encoder-decoder architecture.
• PCSNet generates lung masks from CXR images, which are then fused with original images.
• Classification is performed using an improved lightweight ShuffleNetV2 model on the fused images.

Main Results:

• PCSNet demonstrated high segmentation performance on CXR datasets (MC, SH), outperforming seven other models.
• PCSNet achieved superior accuracy (98.94%) and boundary accuracy (97.86%) compared to Attention-Net with 62% fewer parameters.
• The proposed method improved pneumonia classification accuracy by 0.14% (98.55%) on the CXIP dataset and COVID-19 classification accuracy by 0.1% (97.50%) on the COVIDx dataset, with significant specificity gains.

Conclusions:

• The segmentation-assisted fusion method effectively improves CXR classification accuracy.
• PCSNet offers a computationally efficient and highly accurate solution for lung disease diagnosis from medical images.
• The proposed approach demonstrates clinically meaningful improvements over state-of-the-art methods in medical image analysis.