Hybrid Thompson-UCB for Risk-Aware Classification of Brain Tumours under Limited Data

Brain tumour classification under limited MRI data is not just an accuracy problem. It is a risk-aware task because tumour cases misclassified as healthy may lead to false reassurance. This study reframes four-class brain image classification (glioma, meningioma, pituitary, and healthy) as a single-step contextual bandit problem. We propose a framework based on contextual bandits for cost-sensitive classification framework for risk-aware learning. The framework incorporates cost-sensitive learning directly into the reward system, enabling the agent to focus on clinically risky tumour-to-healthy errors and to reduce them by iteratively improving the classification policy. This paper employs hybrid Thompson Sampling with Upper Confidence Bound (TS-UCB), $\epsilon$-greedy, and Boltzmann exploration policies for risk-aware classification with limited brain-tumour data, and evaluates these policies across three data regimes using cross-validation. The hybrid TS-UCB exploration policy achieves the strongest safety-performance balance, obtaining the lowest tumour-to-healthy error rate and the highest healthy-class precision while maintaining competitive validation and test accuracies. The results further demonstrate that reward shaping can tune the safety-accuracy trade-off by prioritising clinically risky confusions while remaining competitive with, and, in some cases, outperforming traditional deep learning techniques under limited data. These results suggest that hybrid TS-UCB is suitable for risk-aware brain tumour classification when labelled data are scarce and false negatives must be controlled.