Deep learning for screening of interstitial lung disease patterns in high-resolution CT images

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Interstitial lung disease (ILD) diagnosis relies on expert interpretation of high-resolution computed tomography (HRCT) scans.
• Accurate identification of ILD patterns is crucial for prognosis and treatment planning.
• Automated tools could improve efficiency and consistency in ILD detection.

Purpose of the Study:

• To develop and evaluate a deep learning-based screening tool for detecting various interstitial lung disease (ILD) patterns.
• To assess the performance of a fully convolutional network for semantic segmentation of ILD features in HRCT images.

Main Methods:

• A fully convolutional network employing multi-scale feature extraction and dilated convolution was utilized for semantic segmentation of ILD patterns.
• The deep learning model was trained and validated on both a public (MedGIFT) and a private clinical research database.
• Quantitative evaluation was performed using metrics including success rate, sensitivity, and false positives per section.

Main Results:

• The deep learning model demonstrated comparable success rates and sensitivity in detecting fibrosis and emphysema patterns across both databases.
• Detection performance for consolidation patterns was observed to be lower compared to fibrosis and emphysema.
• The study successfully implemented automatic identification of ILD patterns within HRCT images.

Conclusions:

• A deep learning framework provides an effective method for the automatic identification of interstitial lung disease patterns in HRCT images.
• Transfer learning, using a pre-trained model on natural images followed by fine-tuning on a specific ILD database, yielded acceptable results.
• This AI-driven approach shows potential as a screening tool to aid radiologists in ILD detection.