Optimizing Interstitial Photodynamic Therapy Planning With Reinforcement Learning-Based Diffuser Placement

Interstitial photodynamic therapy (iPDT) has shown promising results recently as a minimally invasive stand-alone or intra-operative cancer treatment. The development of non-toxic photosensitizing drugs with improved target selectivity has increased its efficacy. However, personalized treatment planning that determines the number of photon emitters, their positions and their input powers while taking into account tissue anatomy and treatment response is still lacking to further improve outcomes. Objective: To develop new algorithms that generate high-quality plans by optimizing over the light source positions, along with their powers, to minimize the damage to organs-at-risk while eradicating the tumor. The optimization algorithms should also accurately model the physics of light propagation through the use of Monte-Carlo simulators. Methods: We use simulated-annealing as a baseline algorithm to place the sources. We propose different source perturbations that are likely to provide better outcomes and study their impact. To minimize the number of moves attempted (and effectively runtime) without degrading result quality, we use a reinforcement learning-based method to decide which perturbation strategy to perform in each iteration. We simulate our algorithm on virtual brain tumors modeling real glioblastoma multiforme cases, assuming a 5-ALA PpIX induced photosensitizer that is activated at 635 nm wavelength. Results: The algorithm generates plans that achieve an average of 46% less damage to organs-as-risk compared to the manual placement used in current clinical studies. Significance: Having a general and high-quality planning system makes iPDT more effective and applicable to a wider variety of oncological indications. This paves the way for more clinical trials.