Landmark Detection in 3D Medical Images Using Reinforcement Learning

Computer vision tasks like key point localization for face recognition and pose estimation uses Deep Learning (DL) and Reinforcement Learning (RL). Similarly for medical image analysis, anatomical landmark detection is beneficial for automatic scan planning which reduces technician's time and error. This paper compares DL and RL models and variants of Q-learning methods like Deep Q-Network in a single and multi-agent setting. Results show that RL approaches are performing significantly better as compared to DL approaches. All the metrics are in mm. For five landmarks, with single-agent RL, Mean absolute error and Mean euclidean error are 1.66 mm and 3.87 mm, whereas, with multi-agent RL, errors are 2.12 mm and 4.33 mm respectively on the test set (n=60). While with DL approach, errors are 6.54 mm and 14.62 mm which is nearly 3 to 4 times more than RL based approach. The performance of all variants of the DQN approach is excellent, where Double DQN is performing best with errors as 2.10 mm and 4.34 mm.