Figure 4.

Schematic diagram of the proposed 3D cephalometric landmark detection framework using deep reinforcement learning (DRL). The training stage included data import and volume rendering in the first phase, anatomical image view adjustment for landmarks in the second phase, and training with DRL agents in the third phase. The latter is depicted in detail in the upper box with a drawing to illustrate the agent navigating toward the target landmark. The inferencing stage included both single- and multi-stage DRL, starting with the same 3D model rendering in phase 1, followed by single-stage or multi-stage DRL in the second phase, and finalized by gradient-based boundary estimation for single-stage DRL, or by repeated DRL and landmark prediction for multi-stage DRL in the third phase.