Fusing learned representations from Riesz Filters and Deep CNN for lung tissue classification.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Accurate classification of lung tissue in Computed Tomography (CT) is crucial for diagnosing interstitial lung diseases (ILD).
• Existing methods often struggle with variations in tissue micro-architecture and rotation.
• Deep learning models, while powerful, may lack explicit invariance to local image rotations.

Purpose of the Study:

• To develop and evaluate a novel method for detecting and classifying diseased and healthy lung tissues in CT images.
• To investigate the complementary strengths of Riesz and deep learning features for lung tissue analysis.
• To compare early and late feature fusion strategies for improved classification performance.

Main Methods:

• Utilizing Riesz representations to learn steerable, rotation-invariant texture signatures for lung tissue classes.
• Employing deep Convolutional Neural Networks (CNNs), specifically Inception V3, fine-tuned on an augmented ILD dataset for feature extraction.
• Fusing Riesz and deep CNN features within a joint softmax model for final classification, comparing early and late fusion approaches.

Main Results:

• The fused Riesz and deep CNN features significantly improved classification performance compared to individual feature sets.
• Late fusion of independent probabilities demonstrated superior results over early fusion and ensemble deep learning methods.
• The proposed method effectively classifies four diseased tissue types and healthy lung tissue from the ILD dataset.

Conclusions:

• Fusion of Riesz and deep CNN features offers a robust and effective approach for lung tissue classification in CT imaging.
• Late feature fusion is a promising strategy for enhancing diagnostic accuracy in medical image analysis.
• This novel method holds potential for improving the computer-aided diagnosis of interstitial lung diseases.