Synopsys Cloud
Cloud native EDA tools & pre-optimized hardware platforms
Request a Free Trial →
Simpleware Software Used for Comparing Medical Image Segmentation Algorithm Accuracy for 3D Printing
Posted on 30 June 2022 by Kerim Genc
Point of care (POC) 3D printing is growing in popularity as an option for clinicians to generate patient-specific printed models and devices from scan data. We provide a route from 3D images to models for medical 3D printing through our FDA 510(k)-cleared solution, and we were pleased to see Simpleware software included in a recent publication co-authored by researchers at the US Food and Drug Administration (the FDA) and US Department of Veteran Affairs (VA) comparing algorithms used in medical image segmentation.
The publication reviews common algorithms used for patient-specific image data segmentation and mesh refinement, and identifies parameters for introducing geometric variability and measuring variability between models and validating processes. From this approach, methods for understanding the basic segmentation and refinement functions of software can help sites create a baseline to evaluate standard workflows, improve user training, and reduce inter-user variability when using patient-specific models for clinical decision-making.
The paper looks at an image-to-mesh workflow for 3D printing that begins with volumetric clinical imaging such as CT and MRI to capture anatomies. From this point, software segmentation assigns voxels that contain an anatomic region of interest (ROI) and then creates a surface mesh that can be adjusted depending on computational needs and clinical requirements. The authors used an example dataset, chosen for its segmentation complexity, that comprises a cranio-maxillofacial CT scan including a large right mandibular tumor adjacent to a nerve, and an impacted wisdom tooth. As well as Simpleware software, Dicom2Print (3D Systems), Mimics (Materialise), and open-source program 3D Slicer (Brigham and Women’s Hospital) were used to carry out segmentation and create STLs for comparison.
The basic workflow for evaluating the different algorithms involved a single user:
- Importing DICOM data to each software program
- Carrying out segmentation using a mix of automated and semi-automated methods, such as thresholding for bone volume segmentation
- Manually segmenting the nerves by painting on the mask on each visible layer, and then using interpolation (where available) to connect the mask layers
- Generating surface meshes, including with different smoothing and mesh refinement options
- Exporting to STL (or other 3D printer-compatible file formats)
User intervention level at different workflow steps, by program. Automatic designates use of a menu option or built in automatic “turn key” feature for the majority of the segmentation step. Semi-Automatic designates segmentation steps that required user intervention but made use of built-in operation controls where parameters must be set by the user. Manual designates steps that required all or almost all manual user intervention to segment. Final Models after export were used for subsequent calculations (Table by Fogarasi et al. / CC BY 4.0 / Resized from original).
The resulting 3D print-ready STL files were then compared against several metrics. Due to being unable to perform ground truth verification against the original patient, several methods were evaluated to measure the agreement and disagreement between segmented regions of interest:
- Volume and surface area measurements provided global comparisons between models from each software
- Local differences were evaluated through linear measurements for specific ROIs (mandible, tumor, nerve), including comparative analysis between mandible and tumor STL model surface area, STL model volumes, and nerve path centroids
Residual volume comparison from the literature was also employed to calculate tumor volumes for models, and to calculate the agreement and disagreement volume between them. Agreement and disagreement metrics were employed to compare the models, while surface deviation heatmaps were used to identify local areas where variation was more prominent for each model.
Results
Across the different software programs, the time to process the image data was similar but the level of manual intervention needed varied between them. Initial visual inspection showed some differences in segmentation, while there were no statistically significant differences between models. Although some statistical differences were identified, they did not meet the threshold for clinical relevance based on the criteria for variance in tumor morphology and margins, and the location of the nerve did not change in such a way that would be likely to modify a surgical plan. However, the authors emphasize the need for expert manual segmentation of complex anatomical regions, and clinical review of measurements and accuracy, when acting on the models.
Volume representations of mandible segmentations with accompanying mask previews. Note the differences in segmentation of the interior wall of the mandible and location of the nerve in each cross-section. Models presented are after Smoothing 1 modifications were applied per program (Image by Fogarasi et al. / CC BY 4.0 / Resized from original).
Discussion
The authors demonstrated the value of being able to identify and quantify sources of variation between different segmentation and meshing algorithms when ground truth measurements cannot be made for internal patient anatomies, and cadavers and polymer models offer limited detail. However, the study was unable to compare the magnitude of anatomic variations with inter-user or intra-user variability between different scanning protocols or clinical applications.
Using software that already has FDA clearance for segmentation and creation of STL files, such as Simpleware software, is discussed as helping to speed up workflow validation compared to more flexible open-source software requiring in-house or third-party testing. From our perspective, this publication provides excellent guidance to any group currently using or thinking of implementing POC 3D Printing within their institution, especially when considering different algorithm implementations and understanding the variation that can occur between the workflows of any segmentation software.
Learn More
- Read the full article Algorithms used in medical image segmentation for 3D printing and how to understand and quantify their performance published in 3D Printing in Medicine
- Except where otherwise noted, content on this site is licensed under a Creative Commons Attribution 4.0 International license
- Read about Simpleware software for Point of Care 3D Printing
Any Questions?
If you would like to know more, please contact us. Our technical specialists will be happy to show you Simpleware software solutions for any anatomy.