Interpretative computer-aided lung cancer diagnosis: From radiology analysis to malignancy evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computer-aided diagnosis (CAD) systems enhance diagnostic accuracy and reduce radiologist workload.
• Current deep learning models for pulmonary nodule detection are clinically viable, but malignancy evaluation lacks clinical insight.
• Existing methods rely on heuristic inference from low-dose computed tomography (LDCT) scans, missing crucial clinical context.

Purpose of the Study:

• To develop a deep learning model that integrates radiological analysis for accurate pulmonary nodule malignancy evaluation.
• To enhance the clinical cognition of AI-driven diagnostic tools in oncology.
• To provide interpretable insights into the decision-making process of AI models for nodule malignancy assessment.

Main Methods:

• Proposed R2MNet, a network for joint radiology analysis and malignancy evaluation.
• Extracted radiological features as channel descriptors to emphasize critical regions for malignancy assessment.
• Introduced channel-dependent activation mapping (CDAM) for visualizing features and explaining deep neural network (DNN) decisions.

Main Results:

• Achieved Area Under Curve (AUC) of 96.27% for nodule radiology analysis and 97.52% for malignancy evaluation on the LIDC-IDRI dataset.
• CDAM analysis identified nodule shape and density as key factors in malignancy determination.
• The model's decision-making process aligns with the diagnostic reasoning of experienced radiologists.

Conclusions:

• The R2MNet system integrates radiological analysis with malignancy evaluation, improving diagnostic confidence and aligning with radiologist workflows.
• CDAM provides valuable model interpretability, highlighting regions critical for DNN-based malignancy probability estimation.
• This approach enhances the clinical applicability of AI in pulmonary nodule diagnosis.