Multi atlas based segmentation approach for neurosurgery planning and guidance

Neurosurgical plannining and guidance improve treatment effects while minimizing the procedure invasiveness. In this context, the accurate identification of brain’s anatomical structures is a fundamental step for succesful surgery outcome. However, manual tissue identification is a time consuming process, not compatible with the clinical routine. Moreover, it suffers from inter- and intra-subjects variability, thus making weak the entire process. In this work, we tested a new multi atlas based segmentation algorithm for neurosurgery applications. The algorithm is part of an open source software, Plastimatch, featuring registration and label fusion capabilities. The procedure was tested on 8 brain’s structures (both left and right putamen, thalamus, hippocampus and caudate) of 20 healthy subjets. Segmentation quality was evaluated in terms of Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) between surface meshes. The ground truth was represented by manually drawn contours. The median value of DSC was generally higher than 0.80 (with the exception of hippocampus) and HD was lower than 2 mm. The inter-quartile variability for DSC and HD ranged respectively from 0.73 to 0.95 and from 0.90 mm to 2.24 mm. In the case of hippocampus the median±quartiles for DSC and HD were 0.78±0.04 and 1.83±0.37 mm respectively. Finally, the computation time was on average 32 minutes, allowing a feasible clinical usage. In conclusion, the proposed methodology is proved to be transferable into the clinical routine, in order to segment anatomical structures for neurosurgery planning and guidance.