Diffusion Models vs. DCGANs for Class-Imbalanced Lung Cancer CT Classification: A Comparative Study
Effective lung cancer detection from CT scans remains critically challenged by class imbalance where benign and normal cases are underrepresented, leading to biased machine learning models with reduced sensitivity for minority classes and potentially missed diagnoses in cancer screening applications. We present a comprehensive comparative analysis of Diffusion Models and Deep Convolutional Generative Adversarial Networks (DCGANs), both incorporating modern architectural enhancements including spectral normalization, self-attention mechanisms, and conditional generation, for addressing class imbalance in lung cancer CT classification. Using the IQ-OTH/NCCD dataset comprising 1,097 CT images across normal, benign, and malignant categories with statistical validation across 10 independent runs, we evaluated both approaches through quantitative image quality metrics (Frechet Inception Distance, Kullback-Leibler divergence, Kernel Inception Distance, and Inception Score) and downstream classification performance. While Diffusion models consistently outperformed DCGANs across most image quality measures, the clinical significance was confirmed through task-based validation. Both generative approaches successfully addressed class imbalance: DCGAN-augmented datasets achieved overall accuracy of 0.9760 {+/-} 0.0116 with benign recall improvement from 0.833 to 0.933, while Diffusion-augmented datasets reached superior performance of 0.9959 {+/-} 0.0068 with perfect benign recall (1.000 {+/-} 0.000). Critically for cancer screening where false negatives carry severe consequences, Diffusion maintained the highest malignant detection sensitivity (0.997 {+/-} 0.008) with substantially lower performance variance, demonstrating more consistent synthetic data quality. These findings establish that while both modern architectures can mitigate class imbalance, Diffusion models superior recall performance and lower variability position them as the preferred approach for high-stakes clinical applications, demonstrating that ultimate validation must prioritize downstream clinical task performance over image quality metrics alone.