Active vision for brain tumor localization on grid-discretized MRI using a dual-map actor-critic network with proximal policy optimization

Dense expert annotation of medical images remains expensive, and learning reliable localization models from limited clinical datasets is challenging. This work reformulates brain tumor localization as an oracle-terminated active search problem on a discretized 6×6 MRI grid rather than a conventional static prediction task. A Proximal Policy Optimization (PPO) agent is trained to sequentially navigate the grid using a Dual-Map Actor–Critic architecture that fuses high-resolution local texture, downsampled global anatomical context, and normalized spatial coordinates. A dense Manhattan-distance reward guides exploration toward the target region, while oracle termination occurs when the agent reaches a tumor-containing grid cell. Experiments were conducted using a patient-disjoint subset of the Figshare Brain Tumor Dataset comprising 200 training MRI slices and 50 held-out test slices. Under the predefined evaluation protocol, the proposed policy achieved a 96.0% oracle-terminated search success rate (95% confidence interval: 86.3–99.5%) with average trajectory length of 8.26 steps. Because the study evaluates a sequential navigation policy under oracle termination, comparisons with supervised CNN-, U-Net-, and V-Net-based methods are presented only as formulation-specific reference experiments rather than direct performance rankings. These results demonstrate the feasibility of active sequential search for tumor localization under a constrained proof-of-concept setting.