Outlier Detection for Shape Model Fitting

Rahbani, Dana. Outlier Detection for Shape Model Fitting. 2021, Doctoral Thesis, University of Basel, Faculty of Science.


Official URL: https://edoc.unibas.ch/87811/

Downloads: Statistics Overview


Medical image analysis applications often benefit from having a statistical shape model in the background. Statistical shape models are generative models which can generate shapes from the same family and assign a likelihood to the generated shape. In an Analysis-by-synthesis approach to medical image analysis, the target shape to be segmented, registered or completed must first be reconstructed by the statistical shape model. Shape models accomplish this by either acting as regression models, used to obtain the reconstruction, or as regularizers, used to limit the space of possible reconstructions. However, the accuracy of these models is not guaranteed for targets that lie out of the modeled distribution of the statistical shape model. Targets with pathologies are an example of out-of-distribution data. The target shape to be reconstructed has deformations caused by pathologies that do not exist on the healthy data used to build the model. Added and missing regions may lead to false correspondences, which act as outliers and influence the reconstruction result. Robust fitting is necessary to decrease the influence of outliers on the fitting solution, but often comes at the cost of decreased accuracy in the inlier region. Robust techniques often presuppose knowledge of outlier characteristics to build a robust cost function or knowledge of the correct regressed function to filter the outliers.
This thesis proposes strategies to obtain the outliers and reconstruction simultaneously without previous knowledge about either. The assumptions are that a statistical shape model that represents the healthy variations of the target organ is available, and that some landmarks on the model reference that annotate locations with correspondence to the target exist. The first strategy uses an EM-like algorithm to obtain the sampling posterior. This is a global reconstruction approach that requires classical noise assumptions on the outlier distribution. The second strategy uses Bayesian optimization to infer the closed-form predictive posterior distribution and estimate a label map of the outliers. The underlying regression model is a Gaussian Process Morphable Model (GPMM). To make the reconstruction obtained through Bayesian optimization robust, a novel acquisition function is proposed. The acquisition function uses the posterior and predictive posterior distributions to avoid choosing outliers as next query points. The algorithms give as outputs a label map and a a posterior distribution that can be used to choose the most likely reconstruction. To obtain the label map, the first strategy uses Bayesian classification to separate inliers and outliers, while the second strategy annotates all query points as inliers and unused model vertices as outliers. The proposed solutions are compared to the literature, evaluated through their sensitivity and breakdown points, and tested on publicly available datasets and in-house clinical examples.
The thesis contributes to shape model fitting to pathological targets by showing that:
- performing accurate inlier reconstruction and outlier detection is possible without case-specific manual thresholds or input label maps, through the use of outlier detection.
- outlier detection makes the algorithms agnostic to pathology type i.e. the algorithms are suitable for both sparse and grouped outliers which appear as holes and bumps, the severity of which influences the results.
- using the GPMM-based sequential Bayesian optimization approach, the closed-form predictive posterior distribution can be obtained despite the presence of outliers, because the Gaussian noise assumption is valid for the query points.
- using sequential Bayesian optimization instead of traditional optimization for shape model fitting brings forth several advantages that had not been previously explored. Fitting can be driven by different reconstruction goals such as speed, location-dependent accuracy, or robustness.
- defining pathologies as outliers opens the door for general pathology segmentation solutions for medical data. Segmentation algorithms do not need to be dependent on imaging modality, target pathology type, or training datasets for pathology labeling.
The thesis highlights the importance of outlier-based definitions of pathologies in medical data that are independent of pathology type and imaging modality. Developing such standards would not only simplify the comparison of different pathology segmentation algorithms on unlabeled datsets, but also push forward standard algorithms that are able to deal with general pathologies instead of data-driven definitions of pathologies. This comes with theoretical as well as clinical advantages. Practical applications are shown on shape reconstruction and labeling tasks. Publicly-available challenge datasets are used, one for cranium implant reconstruction, one for kidney tumor detection, and one for liver shape reconstruction. Further clinical applications are shown on in-house examples of a femur and mandible with artifacts and missing parts. The results focus on shape modeling but can be extended in future work to include intensity information and inner volume pathologies.
Advisors:Vetter, Thomas
Committee Members:Roth, Volker and Zachow, Stefan
Faculties and Departments:05 Faculty of Science > Departement Mathematik und Informatik > Ehemalige Einheiten Mathematik & Informatik > Computergraphik Bilderkennung (Vetter)
UniBasel Contributors:Vetter, Thomas and Roth, Volker
Item Type:Thesis
Thesis Subtype:Doctoral Thesis
Thesis no:14643
Thesis status:Complete
Number of Pages:xi, 118
Identification Number:
  • urn: urn:nbn:ch:bel-bau-diss146430
edoc DOI:
Last Modified:18 Mar 2022 05:30
Deposited On:17 Mar 2022 12:30

Repository Staff Only: item control page