Imitation learning with deep reinforcement learning improves MRI-guided transurethral ultrasound ablation targeting in simulations and phantoms

Simulation study: an imitation-learning plus deep reinforcement learning controller kept >97% of the ablation boundary within ±1 mm of the target versus <80% for conventional binary and PID controllers; whole-gland precision 88% vs 82%.

• Whole-gland BC + PPO2 was statistically comparable to the conventional approach in precision (85.5%, p=0.31) but reduced treatment time by about 30%, favoring efficiency rather than accuracy.
• The controller autonomously optimizes ultrasound power and rotational speed in real time, while frequency is selected by a predefined rule based on distance to the target boundary.
• Tissue-phantom validation showed consistent performance; the authors note that realistic MRI-condition and in vivo validation are still needed before clinical translation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.