Quantitative Validation of a Knee Segmentation Model on MRI Data

Sammanfattning

At Episurf Medical AB, individualized joint implants are produced. One step to produce the implants is to visualize a 3D model of the distal femur. For this purpose, a deep learning segmentation model is used on MRI images. The model can be used with user input by editing a threshold, and without user input by using a preset threshold. The current segmentation model used was validated using a k-fold cross-validation technique, which is deemed insufficient and needs to be validated on completely unseen data for the model. The segmentation model outputs two separate meshes as STL files, one bone mesh and one merge mesh, thus, methods were developed to compare STL files. Two already established techniques were first used, the Python library Trimesh and the open-source software MeshLab, to generate results from reliable software. Two algorithms were then developed using two PyVista functions, voxelize_volume and select_enclosed_points, to create a voxelized representation that was iterable using Python. The two developed algorithms showed similar results to the established methods. The previous validation showed a bone DSC value of 0.98 ± 0.005 and cartilage DSC value of 0.858 ± 0.028. The bone DSC values for the voxelize_volume algorithm achieved 0.987±0.0062 and 0.981±0.0054 with the edited threshold and non-edited threshold, respectively. The corresponding values for the select_enclosed_points algorithm was 0.987 ± 0.0065 and 0.981 ± 0.0056. The results show that the model is robust towards unseen data. The only comparable result to the previous validation is the bone DSC value with a non- edited threshold, which shows very similar values. The model shows some vulnerability towards artifacts, as the lowest value in most categories was seen in a case with folding artifacts. More cases with clear artifacts are, however, needed to evaluate how robust the model is towards artifacts in general. User input is not always valuable, and for some metrics performs worse with user input. Thus, no clear pattern can be seen, however, it should be considered on a case-by-case basis if it facilitates the workflow.