Unsupervised segmentation of MRI knees using image partition forests

Abstract

Nowadays many people are affected by arthritis, a condition of the joints with limited prevention measures, but with various options of treatment the most radical of which is surgical. In order for surgery to be successful, it can make use of careful analysis of patient-based models generated from medical images, usually by manual segmentation. In this work we show how to automate the segmentation of a crucial and complex joint -the knee. To achieve this goal we rely on our novel way of representing a 3D voxel volume as a hierarchical structure of partitions which we have named Image Partition Forest (IPF). The IPF contains several partition layers of increasing coarseness, with partitions nested across layers in the form of adjacency graphs. On the basis of a set of properties (size, mean intensity, coordinates) of each node in the IPF we classify nodes into different features. Values indicating whether or not any particular node belongs to the femur or tibia are assigned through node filtering and node-based region growing. So far we have evaluated our method on 15 MRI knee images. Our unsupervised segmentation compared against a hand-segmented gold standard has achieved an average Dice similarity coefficient of 0.95 for femur and 0.93 for tibia, and an average symmetric surface distance of 0.98 mm for femur and 0.73 mm for tibia. The paper also discusses ways to introduce stricter morphological and spatial conditioning in the bone labelling process.

Segmentation of hip cartilage in compositional magnetic resonance imaging: A fast, accurate, reproducible, and clinically viable semi‐automated methodology

“…Nor do we believe it would significantly increase the processing time of the semi‐automated method. The addition of these features is something we intend to establish as part of our methodology in further work to optimize our semi‐automated segmentation technique . This would allow the biochemical and morphological analysis of specific areas of cartilage between patients and within patients over time.…”

“…The semi‐automated segmentation method consists of a two‐step process: (i) fully automated hierarchical partitioning of the data volume generated using a bespoke segmentation approach applied recursively, followed by (ii) user selection of the anatomical structure of interest using a bespoke region editor.…”