Semi-supervised multi-class pneumonia classification using a CNN-cascade forest framework

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Disease Diagnosis
• Computational Pathology

Background:

• Accurate medical imaging diagnosis of pneumonia subtypes is crucial but challenged by limited annotated data and imaging variability.
• Current deep learning methods are often univariate and binary-classification, limiting clinical applicability for multi-class pneumonia detection.

Purpose of the Study:

• To propose a semi-supervised CNN-Enhanced Cascade Forest (CE-Cascade) scheme for robust multi-class pneumonia classification using chest X-ray and CT images.
• To address limitations of existing methods by leveraging both labeled and unlabeled data for improved generalization.

Main Methods:

• A convolutional neural network (CNN) extracts deep features from medical images.
• A cascade forest refines these features to identify hierarchical and multi-scale patterns significant for pneumonia detection.
• A semi-supervised pseudo-labeling strategy utilizes unlabeled data to enhance model generalization.

Main Results:

• The CE-Cascade framework was evaluated on 4578 chest X-ray and CT images across bacterial, viral, fungal, general pneumonia, and normal categories.
• The model achieved an overall classification accuracy of 98.86%, surpassing state-of-the-art deep learning systems.
• Experimental findings demonstrate the model's effectiveness in classifying diverse pneumonia types.

Conclusions:

• The integration of CNN and Cascade Forest with semi-supervised learning offers a robust method for automated multi-class pneumonia classification.
• The CE-Cascade approach provides a clinically meaningful solution, overcoming limitations of scarce annotations and cross-imaging variability.
• This automated system shows significant potential for improving diagnostic accuracy in medical imaging.