Automatic Scoring of Multiple Semantic Attributes With Multi-Task Feature Leverage: A Study on Pulmonary Nodules in...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Computer-aided diagnosis (CAD) performance is limited by the gap between computational and semantic features in clinical applications.
• Accurate description of semantic features for lung nodules in CT images is crucial for diagnosis and management.
• Existing methods struggle to effectively bridge the gap between extracted computational features and clinically relevant semantic descriptions.

Purpose of the Study:

• To bridge the gap between computational and semantic features for improved computer-aided diagnosis of lung nodules.
• To leverage multi-task learning (MTL) to explore relationships among semantic features and map heterogeneous computational features to radiologist ratings.
• To develop a semantic attribute scoring scheme for richer quantitative assessment of lung nodules.

Main Methods:

• Employed three multi-task learning (MTL) schemes to integrate deep learning (stacked denoising autoencoder, convolutional neural network) and hand-crafted features (Haar-like, HoG).
• Utilized the Lung Image Database Consortium (LIDC) dataset with 9 semantic features rated by 12 radiologists.
• Trained MTL models to predict semantic feature scores, mapping computational features to radiologist ratings using cross-validation on 2400 nodules.

Main Results:

• The proposed MTL schemes achieved semantic scores closer to radiologist ratings compared to single-task LASSO and elastic net regression.
• Experimental results demonstrate the effectiveness of MTL in leveraging heterogeneous features for semantic description.
• The method showed improved prediction accuracy for semantic attributes of lung nodules.

Conclusions:

• The developed semantic attribute scoring scheme offers richer quantitative assessments, supporting diagnostic decision-making and patient management.
• The ability to automatically associate medical image content with clinical semantic terms can aid in developing advanced medical search engines.
• This approach enhances the clinical utility of computer-aided diagnosis by improving the interpretation of lung nodules.