Real-Time Break-Through Detection in Craniotomy Using Long Short-Term Memory Networks

An important step before the implantation of an invasive brain-computer interface (BCI) is craniotomy, which provides direct access to the brain. During the craniotomy, the precision of cranial drilling is critical, as any deviation or exces-sive penetration can pose significant risks, potentially leading to severe brain injury. Traditional methods of force signal analysis lack effective feedback in complex situations, leading to a risk of over-drilling and potential brain damage. Here, we proposes a Long Short-Term Memory (LSTM) neural network to detect the break-through of the drill bit in real-time. Thrust force data from five drilling experiments on a fancy rat skull specimen are collected. The force data are then split into 0.2-second intervals corresponding to different stages of the drilling process. An LSTM network is trained with these labeled datasets to detect changes in the drilling stage, thereby determining the break-through point. The trained classicfication network gets a accuracy of 96%. Practical drilling experiments on a fancy rat skull specimen validates the feasibility and accuracy of the break-through detection algorithm. The experimental results demonstrate that the method can achieve an over-penetration distance within 15 μm, which is safe for performing craniotomy in rats. Given that this method ensures safety for the stringent requirements of rat craniotomy, it can be expected to provide even greater safety for human craniotomy procedures.