Evaluating Temperature Scaling Calibration Effectiveness for CNNs under Varying Noise Levels in Brain Tumour Detection

Accurate confidence estimation in deep learning is essential for high-stakes applications like medical imaging, where overconfident misclassifications can lead to serious consequences. This study looks at how Temperature Scaling (TS), a method used after training, can improve the reliability of CNNs in brain tumor classification. We create a custom CNN and train it on a dataset of brain MRIs. To simulate real-world uncertainty, we add five types of image noise to the test set: Gaussian, Poisson, Salt & Pepper, Speckle, and Uniform. We evaluate model performance using precision, recall, F1-score, accuracy, negative log-likelihood (NLL), and expected calibration error (ECE) both before and after calibration. The results show that TS significantly lowers ECE and NLL under all noise conditions without losing classification accuracy. These findings support the use of TS as an effective and simple approach to improving the decision confidence of medical AI systems, making their outputs more trustworthy in noisy or uncertain environments.