Integrating Deep CNN Features with Classical Statistical Regression for Interpretable and Uncertainty-Aware Image Classification

Deep convolutional neural networks (CNNs) achieve state-of-the-art performance in image classification tasks, particularly in medical imaging. However, their limited interpretability and lack of formal statistical inference restrict their adoption in high-stakes clinical and regulatory settings. This study proposes a hybrid two-stage framework integrating CNN-derived embeddings with classical statistical regression models to produce interpretable prediction systems with valid confidence intervals. Using a publicly available chest X-ray dataset, deep features extracted from a fine-tuned ResNet architecture were incorporated into logistic and mixed-effects regression models. Bootstrap resampling and permutation testing were employed to quantify uncertainty and compare model performance. Results demonstrate that the hybrid framework maintains competitive predictive accuracy while enabling estimation of odds ratios, hypothesis testing, and calibrated probability outputs. This integration bridges modern deep learning with classical statistical methodology and provides a pathway toward transparent and trustworthy AI deployment