Transfer Unit

Project

TrU-1

Morphological Scoring of Disease States — Learning from Large Anatomical Databases

Project Heads

Stefan Zachow, Georg Duda

Project Members

Felix Ambellan (ZIB) 

Project Duration

01.01.2019 – 31.12.2021

Located at

ZIB

Description

Medical diagnostics and therapy planning are based on a conceptual understanding of healthy anatomical structures as well as pathological variations thereof. This project intents to support clinical decision-making processes with objective mathematical methods to distinguish between healthy and pathological anatomical morphology in a shape-driven manner based on Convolutional Neural Networks (CNNs) and Statistical Shape Models (SSMs). As application we focus on knee osteoarithritis (OA) due to its high incidence and social relevance.

Typical pathological changes of bony tissue related to knee osteoarthrits ordered with increasing severity from left to right.

The project relies on large epidemiological studies and their MRI imaging cohort data stored in volumious databases. In order to access this data we employ our previously developed highly accurate and fully automated segmentation and 3D reconstruction pipeline for femoral and tibial bone geometry incorporating SSMs and CNNs (Preprint).

Databases under consideration for the project.

To achieve an as accurate as possible shape-based OA classification we follow two complementing lines of work within our project, aiming to profit from advances in both of them:

• Shape representations that allow for Riemannian statistics in order to derive new types of SSMs
• Machine learning mechanisms for (disease) classification that utilize features derived from the newly developed SSMs and/or are augmented by them

Generally speaking, SSMs are developed to capture shape variation within a population of objects in a low dimensional fashion. This is accomplished by mapping inputs to a chosen shape representation and performing statistics within the accompanying space. Depending on the structure of this representation space (and the mapping from shape to representation and back again) the variation is captured differently for different approaches and thus the main modes of variation of a model form plausible or rather implausible shape trajectories. As a consequence the coefficients representing input shapes as linear combination of the modes (shape weights) become more or less meaningful measuring anatomically plausible morphological variation, especially regarding pathological shape changes.

Same input, but different description of variation. Left: Input shapes; center: Main mode of variation (interpolation trajectory) for the standard point distribution model w.r.t. the Euclidean structure of vertex coordinates; right: Main mode of variation (interpolation trajectory) w.r.t. a newly developed SSM comprising a nonlinear structure (Preprint). (Data taken from FAUST dataset)

Aiming for a shape-based OA classification as diagnosis support we utliize several machine learning approaches including, but not exclusively relying on deep learning and CNNs. Among others we employ classic classification methods, e.g. support vector machines, spatial geometric deep learning with local patch operators as well as global patch operators that reduce the problem to the established Euclidean deep learning setting.

Geometric features (coordinates, curvature, SSM shape weights, etc.) are used in a machine learning framework for a shape-based OA classification.

Website @ ZIB

Morphological Scoring of Disease States: Learning from Large Anatomical Databases @ ZIB

Selected Publications

• F. Ambellan, M. Hanik and C. von Tycowicz. Morphomatics: Geometric Morphometrics in Non-Euclidean Shape Spaces, A Publicly Available Python Library, 2021, Download, License.
• F. Ambellan, S. Zachow and C. von Tycowicz. Rigid Motion Invariant Statistical Shape Modeling based on Discrete Fundamental Forms, In Medical Image Analyisis, no. 102178, Vol.73, 2021 (Preprint).
• A. Tack, F. Ambellan and S. Zachow. Towards Novel Osteoarthritis Biomarkers: Multi-Criteria Evaluation of 46,996 Segmented Knee MRI, In Medical Image Analyisis, no. 102178, Vol.73, 2021 (Open Access).
• A Sekuboyina, ME Husseini, F. Ambellan, S. Zachow et al.. VerSe: A Vertebrae Labelling and Segmentation Benchmark for Multi-Detector CT Images, In Medical Image Analyisis, no. 102166, Vol.73, 2021 (Preprint).
• F. Ambellan, S. Zachow and C. von Tycowicz. Geodesic B-Score for Improved Assessment of Knee Osteoarthritis, In Proceedings Information Processing in Medical Imaging (IPMI 2021), pp. 177-188, Vol.12729, Lecture Notes in Computer Science, 2021 (Preprint).
• F. Ambellan, S. Zachow and C. von Tycowicz. A Surface-Theoretic Approach for Statistical Shape Modeling, In Proceedings Medical Image Computing and Computer Assisted Intervention (MICCAI 2019), Part IV, pp. 21-29, Vol.11767, Lecture Notes in Computer Science, 2019 (Preprint).
• F. Ambellan, S. Zachow and C. von Tycowicz. An as-invariant-as-possible GL+(3)-based Statistical Shape Model, In Proceedings Mathematical Foundations of Computational Anatomy (MFCA at MICCAI 2019), pp. 219-228, Vol.11846, Lecture Notes in Computer Science, 2019 (Preprint).
• F. Ambellan, H. Lamecker, C. von Tycowicz, and S. Zachow. Statistical Shape Models – Understanding and Mastering Variation in Anatomy, In Biomedical Visualisation, Paul M. Rea (Ed.), Springer Nature Switzerland AG, 1, pp. 67-84, 2019 (Preprint)

Selected Pictures