Optimizing pulmonary chest x-ray classification with stacked feature ensemble and swin transformer integration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pulmonary chest x-ray analysis is crucial for diagnosing lung conditions.
• Automating this process can enhance diagnostic efficiency and accuracy.
• Current methods may face limitations in feature extraction and generalization.

Purpose of the Study:

• To develop an integrated framework for automated pulmonary chest x-ray image classification.
• To improve the accuracy and efficiency of chest x-ray analysis using advanced deep learning techniques.
• To leverage feature ensemble and optimized transformer architectures for enhanced performance.

Main Methods:

• Utilized a feature ensemble approach combining outputs from pre-trained convolutional neural networks (CNNs) like VGG16, ResNet50, and MobileNetV2.
• Developed 3D image representations from stacked 2D grayscale pooled features for classifier input.
• Incorporated the Swin Transformer architecture, optimized with the Artificial Hummingbird Algorithm (AHA) for hyperparameter tuning.

Main Results:

• Achieved high classification accuracies across three diverse datasets: VinDr-CXR (98.874%), PediCXR (98.528%), and MIMIC-CXR (98.958%).
• Demonstrated the effectiveness of the stacked ensemble technique in creating comprehensive feature representations.
• Showcased the robustness and generalizability of the AHA-optimized Swin Transformer model.

Conclusions:

• The proposed integrated framework offers a highly accurate and efficient solution for automated pulmonary chest x-ray classification.
• The combination of stacked features and AHA-optimized Swin Transformer shows significant potential for clinical applications.
• The model's performance across multiple datasets highlights its reliability in various imaging conditions.