EMPHYSEMA CLASSIFICATION USING A MULTI-VIEW CONVOLUTIONAL NETWORK.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Emphysema classification in Computed Tomography (CT) images is crucial for diagnosing lung diseases.
• Current methods may be complex or less accurate, necessitating simpler, effective automated tools.

Purpose of the Study:

• To propose and validate a fully automatic tool for emphysema classification using a simple Convolutional Neural Network (CNN).
• To compare the performance of the proposed CNN architecture against more complex and deeper state-of-the-art models.

Main Methods:

• A simplified CNN architecture with 4 convolutional and 3 pooling layers was developed.
• The model utilizes a 2.5D multiview representation (axial, sagittal, coronal) of pulmonary tissue.
• The proposed 2.5D CNN was compared against 2D CNN, 3D CNN, and deeper, more complex architectures.

Main Results:

• The proposed method achieved an overall sensitivity of 81.78% and a specificity of 97.34% on 1553 tissue samples.
• The simplified CNN architecture demonstrated superior performance compared to deeper, state-of-the-art lung pattern classification models.
• Satisfactory results were observed for full-lung emphysema classification.

Conclusions:

• A simple CNN architecture can effectively learn discriminative features for emphysema classification from CT images.
• The proposed 2.5D CNN tool offers a highly accurate and efficient alternative to more complex models for emphysema detection.
• This automated approach shows significant potential for clinical application in diagnosing lung conditions.