Ensemble fusion model for improved lung abnormality classification: Leveraging pre-trained models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Lung abnormalities require prompt diagnosis for effective treatment.
• Smartphone-based chest X-rays offer potential for wider accessibility.
• Accurate sub-classification of abnormalities is crucial for targeted therapy.

Purpose of the Study:

• To develop and evaluate a novel deep learning methodology for sub-classifying lung abnormalities from smartphone-captured chest X-rays.
• To improve the accuracy and efficiency of lung abnormality diagnosis using artificial intelligence.
• To provide a viable solution for resource-constrained environments.

Main Methods:

• A Convolutional Neural Network (CNN) with three max pooling layers and early fusion was designed.
• The CheXpert dataset was utilized, with 13 distinct sub-classes of lung abnormalities.
• Dedicated sub-models were trained for each abnormality sub-class, with outputs integrated via early fusion (Method 1) or an ensemble approach (Method 2).

Main Results:

• The proposed 3M-CNN and fused ensemble models achieved a high accuracy of 98.79%.
• This accuracy surpasses existing methodologies for lung abnormality sub-classification.
• The approach demonstrates effectiveness for smartphone-based medical imaging.

Conclusions:

• The novel CNN-based methodology offers a highly accurate and efficient solution for sub-classifying lung abnormalities from smartphone X-rays.
• This technique has significant implications for improving diagnostic capabilities, particularly in resource-limited settings.
• The study highlights the potential of AI-powered mobile health solutions in radiology.