Enhancing Clinical Decision Support through Cost Sensitive CNN and Reliability Calibrated Pneumonia Classification

Pneumonia detection from chest X-ray images is widely used in computer-aided diagnostic systems. However, effective clinical decision support requires not only accurate classification performance but also consideration of unequal error costs, since false negative predictions may lead to more severe consequences than false positives. In addition, prediction probabilities must be well calibrated to support threshold-based medical decisions such as triage and patient escalation. This research investigates asymmetric misclassification costs and probability calibration for binary classification (PNEUMONIA vs. NORMAL) using the Hugging Face dataset hf-vision/chest-xray-pneumonia. The proposed framework utilizes a ResNet-18 architecture integrated with cost-sensitive learning through weighted cross-entropy loss (FN:FP = 5:1), threshold optimization based on validation data to reduce expected cost, and post-hoc temperature scaling for improving probability calibration. Experimental results on the independent test set indicate that the cost-sensitive approach enhances specificity and decreases expected cost compared to the conventional cross-entropy baseline. Furthermore, temperature scaling improves the reliability of probabilistic predictions, as demonstrated by better negative log-likelihood and Brier score values. The study also explores selective prediction strategies to balance prediction coverage and risk reduction, complemented by Grad-CAM visualizations and structured failure-case analysis for qualitative assessment. Overall, the findings demonstrate that incorporating cost-aware decision thresholds and calibrated probability estimates can serve as lightweight yet effective enhancements for chest X-ray classification systems in clinical decision-support applications.