Diffusion models vs. DCGANs for class-imbalanced lung cancer CT classification: A comparative study

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Class imbalance in lung cancer CT scans leads to biased models and missed diagnoses.
• Benign and normal cases are often underrepresented, impacting screening sensitivity.

Purpose of the Study:

• To compare Diffusion Models and Deep Convolutional Generative Adversarial Networks (DCGANs) for addressing class imbalance in lung cancer CT classification.
• To evaluate generative approaches using image quality metrics and downstream classification performance.

Main Methods:

• Comparative analysis of Diffusion Models and DCGANs with spectral normalization, self-attention, and conditional generation.
• Utilized the IQ-OTH/NCCD dataset (1097 CT images) with 10 independent runs for validation.
• Evaluated using Frechet Inception Distance, KL Divergence, Kernel Inception Distance, Inception Score, and classification accuracy.

Main Results:

• Diffusion models showed superior image quality metrics and downstream classification performance compared to DCGANs.
• Both methods improved benign recall; Diffusion models achieved perfect benign recall (1.000 ± 0.000) and higher overall accuracy (0.9959 ± 0.0068).
• Diffusion models demonstrated higher malignant detection sensitivity (0.997 ± 0.008) with lower performance variance.

Conclusions:

• Diffusion models are the preferred approach for high-stakes clinical applications like cancer screening due to superior recall and consistency.
• Downstream clinical task performance is critical for validating generative models, not just image quality metrics.
• Both Diffusion Models and DCGANs can mitigate class imbalance, but Diffusion Models offer enhanced reliability for medical diagnosis.