Uncertainty-aware image classification on 3D CT lung.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Early lung cancer detection is vital for patient survival, with deep learning models showing promise.
• Current models often lack reliability and robustness, exhibiting overconfidence on unseen data.
• Model uncertainty can guide referral to medical experts for critical second opinions.

Purpose of the Study:

• To develop and evaluate an uncertainty-aware framework for classifying benign and malignant lung nodules from 3D CT images.
• To quantify prediction uncertainty using Monte Carlo Dropout (MCD), Deep Ensemble (DE), and Ensemble Monte Carlo Dropout (EMCD).
• To assess the impact of uncertainty quantification and data referral on diagnostic performance.

Main Methods:

• Proposed a three-phase framework: data preprocessing/model selection, uncertainty quantification (UQ), and uncertainty measurement/data referral.
• Evaluated eight deep learning models (ResNet, DenseNet, Inception family), employing MCD, DE, and EMCD for UQ.
• Utilized 3D CT images for nodule classification, comparing UQ approaches and implementing a data referral threshold.

Main Results:

• All evaluated deep learning models achieved average F1 scores above 0.832, with InceptionResNetV2 reaching 0.845.
• Incorporating UQ significantly improved overall model performance.
• MCD excelled in uncertainty estimation, while DE and EMCD demonstrated superior U_Recall, identifying incorrect predictions effectively.
• A data referral threshold further enhanced accuracy to 0.959.

Conclusions:

• The proposed uncertainty-aware framework enhances the reliability and robustness of lung nodule classification from 3D CT scans.
• Uncertainty quantification methods, particularly DE and EMCD for U_Recall, are crucial for identifying unreliable predictions in medical diagnostics.
• Implementing a data referral strategy based on uncertainty thresholds significantly boosts diagnostic accuracy, aiding clinical decision-making.