Learning Latent Spiculated Features for Lung Nodule Characterization.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Machine Learning for Healthcare

Background:

• Computer-aided diagnosis (CAD) systems lack transparency in their decision-making processes.
• Explainability in CAD is crucial for clinical adoption, especially in interpreting visual characteristics of medical images.
• Current machine learning models often provide predictions without clear reasoning.

Purpose of the Study:

• To develop an explainable AI approach for CAD systems that leverages visual characteristics for diagnostic interpretation.
• To encode image features based on similarity rather than solely on content.
• To automate the diagnostic characterization of lung nodules by quantifying spiculation.

Main Methods:

• Utilized a Siamese convolutional neural network (SCNN) to learn similarity among lung nodules.
• Encoded image content using SCNN's similarity-based feature representation.
• Applied the K-nearest neighbor (KNN) algorithm for diagnostic characterization using Siamese-based features.

Main Results:

• Demonstrated the feasibility of quantifying spiculation, a key characteristic for malignancy assessment.
• Successfully applied the method to the NIH/NCI Lung Image Database Consortium (LIDC) dataset.
• Showcased the potential for automated diagnostic characterization of lung nodules.

Conclusions:

• Spiculation in lung nodules can be effectively quantified using explainable AI methods.
• This approach enhances the transparency and clinical utility of CAD systems.
• The developed method automates diagnostic characterization, supporting radiologists in cancer malignancy assessment.