Author: Md Mainul Abrar, Yujie Chi 👨🔬
Affiliation: University of Texas at Arlington, Department of Physics, University of Texas at Arlington 🌍
Purpose: Healthcare 5.0, proposed in 2021, includes interpretable healthcare analysis as a core component. Achieving this requires the application of explainable artificial intelligence (XAI) to overcome the verifiability and resilience limitations of black-box AI approaches. In this study, we demonstrate XAI’s potential in identifying high-performing AI agents, using an actor-critic with experience replay (ACER)-based automatic treatment planner for radiotherapy treatment planning as a testbed.
Methods: We trained an ACER agent for dose-volume histogram (DVH)-based inverse treatment planning in intensity-modulated radiotherapy. In this study, we applied XAI with an integrated gradient-based attribution method to understand how DVH states influence the ACER agent's decision-making. This method was applied to intermediate agents, generating heatmaps that show the impact of each DVH point on the agent's actions. By analyzing the heatmaps, we identified decision-making patterns and selected two agents, Agent-E1 and Agent-E2, based on superior heatmap performance. We then compared their treatment planning effectiveness with Agent-C, chosen for its performance on the convergence map.
Results: XAI heatmaps revealed that the ACER agent gradually learns to respond effectively to DVH-based plan quality criteria as training progresses, despite being blind to the criteria. Agents-C, E1, and E2 were selected from training steps 120,500, 160,000, and 200,000, respectively. Testing on 109 cases with an average initial plan score of 6.1±2.0 (out of 9), Agent-C achieved a score of 8.9±0.3 after 19.1±5.9 planning steps. In comparison, Agent-E1 and Agent-E2 improved the score to 9.0±0.0 and 9.0±0.1, respectively, after 14.6±3.8 and 13.9±3.4 planning steps. The three agents improved 92.7%, 99.1%, and 95.4% of the cases to a full score of 9. This demonstrates that XAI-based agents outperform those selected upon regular convergence maps, achieving higher success rates and quicker convergence.
Conclusion: XAI can help identify high-performing AI agents for automatic treatment planning, thereby improving the verifiability of model predictions.