Abstract | Keywords | Export | Availability | Bookmark

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Pedicle screw placement (PSP) is a surgical procedure in spine fusion surgery to keep the vertebrae fixed. The main risk of PSP is that penetration of the pedicle wall, termed pedicle breach, can lead to vascular, visceral and other complications. A ball tip feeler (BTF) gives haptic feedback to a surgeon when detecting a pedicle breach. By palpating the walls of the pedicle with a BTF, breaches can be detected before screws are placed. However this technique appears to be unreliable even when the surgeon is highly experienced. Some research has been conducted using ultrasound, accelerometer, load cell or optical sensor in the past years. Possible reasons why prototypes with these sensing technologies have not yet been built or tested can be because they are either too expensive, complex, time consuming or not fully automatic. The need for an intelligent way to measure breaches remains but the solution will have to be cheap, fast, easy to use and accurate. IMU and loadcell are two cost friendly choices to detect the breaches by measuring the accelerations, orientation and force of the BTF. Because of the complexity of the data, deep learning is used to find patterns and correlations between the data and detecting breaches. After comparing multiple neural networks (NN) the layer structure found for this application uses a combination of a 1D convolutional and a long short-term memory NN. Subsequently, the data is processed in real time, feeds through the model and gives live feedback to the surgeon by beeping when a breach is detected. To evaluate the system's performance, the developed program is used to detect the breach in a realistic pedicle representation using wooden board and 3D printed tube separately. For experimental results, the testing with a custom designed wooden board got a result of 95% accuracy on a set of 207 test data samples. The same model is not transferable onto the 3D-printed tubes. In this setup the BTF reacts too sensitive and produces more than 50% false positives. After gathering new train data using a standardized method a model was retrained for this more complex setup. Here a model with a high accuracy of 92% on the test data was found. Although the model is trained well on the training data, it could not reproduce the same results on real time new experimental data. Some biases were found in the dataset used for training. The dataset lacked breaches on the side of the sample range, input data to include a start and stop motion and was not trained properly on the change of load cell inputs. Another model with an accuracy of only 75% does work in combination with a standardized method on how to approach, interpret and analyse the experiment. Here an accuracy of 86% correctly predicted 3D printed tubes off a set of 36 was achieved with only 9% false positives. This compared to the 80% false positives predicted in a study about BTF reliability by MD Sedory in 2010 is a big improvement. In conclusion, this paper develops a ball tip feeler with a load cell and an IMU, used as input for a complementary AI based breach detection for PSP. Even, the development of this smart BTF is still in proof-of-concept stage. It has a promising potential to be applied in clinical scenarios as a low-cost and user-friendly surgical instrumentation. For future research, gathering quality data, building forward on the NN, training and testing is key to bring this proof of concept to the next stage.