gms | German Medical Science

Introduction: The development and improvement of segmentation algorithms for computer tomography (CT) scans to isolate single organs is of great interest in medical research. However, little work has been done on efficiently analyzing segmented organs regarding morphological characteristics as well as intra- and inter-patient variability. The focus of this work lies on extraction of consistent features from 3D point clouds of multiple organs and unsupervised clustering to identify abnormal geometry within organs and within patients.

Methods: This work utilizes CT-ORG, an open dataset with 140 CT scans from human patients encompassing pre-segmented labels for various organ classes such as liver, bladder, lungs and kidneys [1]. For each organ, descriptors for shape and size are computed from 3D point clouds using the programming language R [2] including the packages “shapes” [3], “Morpho” and “Rvcg” [4], for instance. The data is then clustered and visualized by multiple co-inertia analysis using the package “omicade4” [5], allowing a multi-organ clustering in the same 2-D plot.

Results: First results demonstrate that some individuals exhibit unusual geometric proportions between multiple organs. E.g., some individuals show a close relationship between their organs while for some individuals single organs cluster further away from the other organs. The analysis also yields a 2-D represenation of the variable space thus supporting to identify variables that contribute most to the clustering. Outliers can also be detected when plotting only 2-D coordinates of individual organs.

Discussion: An emerging field in medical imaging deals with computational shape analysis of organs to predict particular diseases. This includes most notably neurodegenerative disorders by interpreting brain scans (reviewed by [6]) as well as more recent approaches on abdominal organs like liver and spleen to predict diabetes [7] for example. Our work aims to assess morphological interrelations among respiratory and abdominal organs independent from disease status. In our example data, the presented approach was mainly successful to identify abnormal between-organ relationships of individual patients. In other examples, the approach is also supposed to identify subclusters of patients. The co-inertia analysis also depends very much on the input data and the shape information extracted from the 3-D data.

Conclusion: While most clustering approaches for shape data from CT scans only allow to represent clustering of single organs, the co-inertia analysis allows to generate a multi-organ cluster representation in the same plot.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.