Advancing automated tuberculosis detection in chest radiographs through stable policy optimization with clipped objectives.

Automated tuberculosis detection from chest radiographs is a fundamentally hard problem that arises from the imbalanced nature of data and weight of clinical misclassifications. In this work, we propose a novel optimization method named clipped Proximal Policy Optimization (PPO) for cost-sensitive sequential decision processing to enhance the computer-aided TB screening on chest X-rays. We first pretrain deep convolutional neural networks (DenseNet201 or ChexNet) via transfer learning on Z-score normalized 224 × 224 images and evaluate this model on binary TB classification. This is achieved in two steps: first, transforming the baseline classifier into a reinforcement learning (RL) agent for which actions correspond to decisions made by the diagnostician and for which the reward function explicitly penalizes false negatives and referrals that were not necessary. Comparative performance reveals the PPO-clipping improves over standard supervised learning as demonstrated by increased accuracy, F1-score, area under the ROC curve and reduced cumulative misclassification costs across held-out test images. Results of ablation experiments confirm that policy optimization results in more resource-efficient and robust diagnostics, particularly for high-stakes or imbalanced scenarios. Our results provide support for embedding clipped-objective RL as a component of deep radiographic classification pipelines to improve safety and clinical viability of TB triage.