Medical image analysis methods allow measurements to be made on biological structures, either by direct measurements or through structures detection (segmentation and extraction of 3D objects for visualization). Regardless of the field of application or the imaging modality, many automatic algorithms are developed to improve the reliability and accuracy of these measurements. On the other hand, there is a lack of standardization of methods for assessing this reliability and accuracy, even more so when there is no gold standard reference to compare. In addition, the clinical goals behind the image analysis task are often not considered for performance evaluation, and therefore there is no hierarchical information between errors. For example, for 3D reconstructions of the scoliotic spine, an error on a measure which determines the surgical decision for the patient would have more impact than an error on a secondary measure that does not interfere with the surgical decision.
The objective of this work is to participate in improving the evaluation methods in the absence of a gold standard by proposing a reference construction methodology based on field experience, and an evaluation methodology that considers the clinical criteria that allow to qualify the performance achieved by an automatic method. The proposed evaluation approach allows for the inclusion of a methodology for reference with the help of experts that allows for more reliable measurement confidence intervals on images, including discussions between experts. The evaluation method includes these confidence intervals as well as new clinical criteria inspired by dedicated research and medical societies related to our field of application, the analysis of scoliosis, and our imaging modality, 3D reconstructions from two-plane X-rays with the EOS system. The methodology put in place is generalizable to other areas and other imaging modalities where the same problems would be encountered.
This method was subsequently transferred to the partner company to be able to conduct evaluations on automatic 3D spinal reconstruction algorithms for pediatric patients and adult patients with spinal deformities.
| Date | 6 Aug 2021 |
|---|
| Original language | French |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Jacques A. de Guise (Supervisor) & Carlos Vázquez (Co-supervisor) |
|---|
Bonhomme, M. (Author), de Guise (Supervisor) &
Vázquez (Co-supervisor),
6 Aug 2021Student thesis: Master's thesis › Master in Engineering: Engineering