Lung cancer diagnosis using deep attention-based multiple instance learning and radiomics.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early lung cancer diagnosis is critical for effective treatment.
• Current computer-aided diagnosis (CAD) methods often analyze lung nodules in isolation, deviating from clinical practice.
• Limited interpretability of existing CAD outputs hinders clinical adoption.

Purpose of the Study:

• To develop a more clinically relevant and interpretable lung cancer diagnosis system using multiple instance learning (MIL).
• To leverage radiomics features and deep attention-based MIL for improved diagnostic performance.
• To address challenges of small and imbalanced datasets in lung cancer diagnosis.

Main Methods:

• Lung cancer diagnosis framed as a multiple instance learning (MIL) problem.
• Utilized radiomics for input features and a deep attention-based MIL algorithm.
• Introduced a novel bag simulation method to enhance MIL performance on small, imbalanced datasets.

Main Results:

• Achieved a mean accuracy of 0.807 (SEM 0.069) and an AUC of 0.842 (SEM 0.074).
• Demonstrated superior performance compared to other MIL methods.
• The proposed oversampling strategy significantly improved model performance.

Conclusions:

• The developed attention-based MIL method offers a more interpretable and accurate approach to lung cancer diagnosis.
• Radiomic features combined with the MIL framework enhance the explainability of diagnostic results for clinicians and patients.
• The method shows promise for improving early lung cancer detection and treatment planning.