Expert knowledge-infused deep learning for automatic lung nodule detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computer-aided detection (CADe) of pulmonary nodules using computed tomography (CT) is vital for early lung cancer diagnosis.
• Deep learning models, while effective, struggle with feature extraction complexity in CT lung images, especially with limited datasets.
• Engineered features offer a complementary approach to enhance CADe performance.

Purpose of the Study:

• To develop a novel CADe methodology for pulmonary nodules that integrates engineered features into deep learning.
• To mitigate challenges associated with limited datasets and reduce the computational complexity of self-learning in nodule detection.
• To improve the accuracy and efficiency of lung nodule detection in CT images.

Main Methods:

• A novel CADe methodology was proposed, infusing engineered features, particularly texture features, into the deep learning process.
• Convolutional Neural Networks (CNNs) were utilized, with engineered features incorporated to prevent overfitting and reduce self-learning complexity.
• The methodology was validated using the public LIDC-IDRI database, focusing on patients with juxta-pleural nodules.

Main Results:

• The proposed methodology achieved a sensitivity of 88% with 1.9 false positives per scan.
• An improved sensitivity of 94.01% was reached with 4.01 false positives per scan.
• The results demonstrate high performance in accuracy and efficiency.

Conclusions:

• The developed CADe methodology effectively integrates engineered features with deep learning for pulmonary nodule detection.
• The approach shows superior performance compared to existing CNN-based and engineered feature-based classification methods.
• This technique offers a promising advancement for accurate and efficient early lung cancer diagnosis via CT imaging.