An Innovative Learning-Based Framework for Automated Craniotomy Planning in Glioma Resection

Objective: Glioma resection remains one of the most challenging procedures in neurosurgery due to the tumor's high malignancy and prevalence. As a critical step in surgical intervention, craniotomy requires meticulous planning to achieve maximal tumor removal while minimizing neurological damage. However, current automated surgical planning methods face significant limitations in addressing craniotomy design, primarily due to the lack of explicit visual targets (e.g., vascular structures) and standardized geometric constraints for bone flap delineation. In this study, we propose an innovative learning-based framework specifically designed for automated craniotomy planning in glioma resection. Approach: Our approach effectively integrates preoperative imaging data and expert demonstrations into a reinforcement learning (RL) model to determine the optimal bone flap geometry. The key innovations of our method include: (1) a self-supervised learning strategy for implicit quantification of glioma, (2) an encoding method for craniotomy pattern designs, (3) a physics-based simulation engine for craniotomy policy training, and (4) an imitation learning-inspired planner for craniotomy planning. Experimental validation was conducted using a dataset derived from publicly available glioma patient images. Main Results: The proposed method presents a success rate of 92.31%±3.85% when processing known craniotomy parameters, and a success rate of 64.5% ±5.87% in end-to-end craniotomy planning from raw preoperative images to definitive surgical plans. Significance: The results demonstrate that our proposed method achieves human-level performance in craniotomy planning, and shows promising potential for end-to-end craniotomy planning from raw preoperative images to definitive surgical plans. Our research provides a valuable reference for the development of intelligent decision-support tools for future neurosurgical procedures.