Toward Robust Lung Cancer Diagnosis: Integrating Multiple CT Datasets, Curriculum Learning, and Explainable AI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computer-aided diagnostic systems show promise in medical imaging but struggle with generalizability and credibility.
• Physicians and specialists criticize current models due to sensitivity and lack of trust.
• Enhancing the generalizability and understandability of diagnostic models is crucial for clinical adoption.

Purpose of the Study:

• To propose a novel deep learning model for enhanced lung cancer diagnosis.
• To improve the quality, understandability, and generalizability of computer-aided lung cancer detection.
• To address the limitations of current models in terms of bias and subjectivity.

Main Methods:

• Utilized five computed tomography (CT) datasets to ensure diversity and heterogeneity.
• Implemented mixup augmentation to combine features and labels, reducing bias and improving generalization.
• Employed curriculum learning for efficient model training, starting with simpler data.

Main Results:

• Achieved high accuracy (99.38%), precision, specificity, and AUC (100%), with sensitivity (98.76%) and F1-score (99.37%).
• Demonstrated minimal false positive (0%) and false negative (1.23%) rates on internal datasets.
• Validated externally with optimal results (100% accuracy, 0% false positives/negatives) and employed explainable AI (Grad-CAM) for interpretability.

Conclusions:

• Developed a robust and interpretable deep learning model for lung cancer diagnostics.
• The model exhibits improved generalizability and validity, overcoming limitations of existing systems.
• Mixup and curriculum learning, combined with diverse datasets, show promise for clinical diagnostic applications.