| Deep brain stimulation(DBS)is considered to be an effective method for the treatment of advanced Parkinson’s disease.MRI-driven preoperative path planning is desired for DBS before the procedure.The preoperative path planning often relies on experienced neurosurgeons to determine the stimulation target and select a puncture path away from the risk structures.Due to the complicated structure of the human brain,the manual path planning on MRI data is a time-consuming task and heavily relies on the experience of the surgeons.The autonomous preoperative path planning system is expected to overcome the above shortcomings.For an ideal autonomous preoperative path planning system,the following aspects are required:1)The system is efficient in computation,it can not only reduce the doctors’ work intensity,but also save the planning time;2)The surgical planning process needs to consider the detailed information of the patient’s brain anatomy from all aspects,because a large number of brain tissues are small in size and have different shapes;3)Target calculation and risk tissue avoidance should fully consider the differences in the shape and size of brain tissue of different patients.The system can automatically and accurately locate the key tissues in the patient’s brain related to the surgical plan.4)Highly autonomous,it can automatically calculate target points,candidate cranial entry points and automatically plan safe puncture paths based on preoperative medical images of patients.In this dissertation,some key issues of autonomous surgical planning for DBS based on deep learning have been deeply studied.The high-precision segmentation of key brain tissues is achieved,the autonomous calculation of target points and candidate cranial entry point sets is completed,and the safest puncture path is autonomously achieved after applying a cost function for each candidate puncture path.The main contributions of this work are summarized as follows:(1)In order to realize the efficient computing operation for the autonomous surgical planning system,a plug-and-play lightweight convolutional module is proposed to reduce the parameters and the floating-point operation in 3D convolution.The lightweight convolution module makes full use of the nature of a large number of redundant features in the convolution feature maps.Through the orthogonal decomposition method of 3D convolution and the cheap generation method of redundant features,a large number of parameters and floating-point operations are compressed.(2)Brain surgery usually requires detailed spatial information for delicate puncture path planning,however,conventional deep neuron networks can hardly give such a solution due to the compression operation of their pooling modules.To give a delicate features detection capability of a neuron network while maintaining reasonable computation cost,this dissertation proposes a learnable orthogonal pooling method which can achieve orthogonal pooling for the 3D input data and feature tensors in the convolution process.By adopting the orthogonal approach,the pooling portion here can retain detailed features of the 3D input data while reducing the amount of network floating-point calculations to cope with hardware resource constraints.(3)In order to fully consider the differences in the shape and size of the key brain tissues of different patients,so as to accurately locate the key tissues related to the surgical plan in the brain of different patients.a 3D high-resolution multistage optimization segmentation strategy is proposed to solve the problem of low contrast and extreme size imbalance in brain tissue,which leads to the decline of segmentation accuracy.By extending the 2D high-resolution network of natural image processing to the field of 3D medical image processing,the problem of the loss of a large amount of detailed information of small objects caused by the pooling operation is improved.And by replacing the conventional 3D convolution with the lightweight plug-and-play convolution module,the number of network parameters and the amount of floating-point operations are reduced.(4)For the autonomous target point determination,two methods are proposed.The first one is an end-to-end target position prediction method based on the learnable generalized orthogonal pooling network proposed in this dissertation.The learnable orthogonal pooling can retain the three-dimensional anatomical structure information of the dimension to be pooled,so that the output variables after pooling have some capability of specific medical physical properties of the orientation plane,enabling end-to-end target coordinate prediction.The second target position determination method highly mimics the neurosurgeon’s operation behavior based on the segmentation results of brain tissues.Both of the two autonomous methods can accurately predict the target coordinates.For the calculation of candidate cranial entry point sets,an autonomous method based on the operating criteria of the neurosurgery chief expert is proposed.Through the clinical statistics of prior angle constraint relationship between the puncture path and the coordinate plane and the setting of the safe sphere space for the candidate cranial entry point,the effective candidate cranial entry points that avoid the sulcus are successfully screened.The safest puncture path is autonomously achieved after applying a cost function for each candidate puncture path based on the anatomical structure of brain risk tissue.The experimental results show that our method provides surgeons with automatic and accurate electrode implant trajectory comparable or even better than manual planning of fellow surgeons. |
