Elsevier

Applied Mathematics and Computation

Volume 267, 15 September 2015, Pages 456-464
Applied Mathematics and Computation

Conformance criteria for validation of target volume surface reconstructed from delineation

https://doi.org/10.1016/j.amc.2015.01.105Get rights and content

Abstract

This paper presents two conformance indexes that can be used to measure a similarity between two delineated structures. The authors use the presented conformance indexes for rating the results of Poisson Surface Reconstruction algorithm on real patient datasets. The rating is based on three delineations: original contours created during a standard procedure, reconstructed contours and reference contours created with an extra care by another physician.

The chosen conformance indexes compare the delineations in two different forms: as a voxel grid and as a surface mesh. In this paper each of the conformance indexes is calculated in three different modes: the standard mode where all differences between datasets are taken into account and two more measuring how much one dataset exceeds the other. The last two modes are not symmetrical.

The proposed conformance indexes allowed us to compare efficiently two delineations. The presented results also allow to state that the use of mesh reconstruction algorithms can improve delineations prepared within a limited time frame.

Introduction

The paper tackles the problem of medical data processing. It describes new ways of comparing two different delineations of the same structure (an organ or a tissue). Delineation is a set of 2D contours marking a volume depicted using an imaging technique like Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) scans. The contours are frequently stored as polygons in XY slices.

In this article we use the term conformance index as a measure of similarity between two delineations. The need for good conformance indexes emerged from the research aimed at improving of the process of radiotherapy planning. With help of Poisson Surface Reconstruction algorithm for automatic generation of Regions of the Interest (ROIs) an adaptive radiotherapy could be implemented.

In our experiment we evaluate our proposed conformance indexes with real patients’ DICOM (Digital Imaging and Communications in Medicine) datasets obtained during the process of radiotherapy planning. A sample CT image with a marked delineation is presented in Fig. 1.

Our goal is to propose a conformance index useful for measuring the similarity of two datasets. We use two different approaches. Although they seem to be similar in general, they take different aspects of the geometry of the provided delineations into account. One conformance index operates on voxel grids with the resolution of the original DICOM dataset, whereas the second index compares 3D surface meshes with an arbitrary resolution.

Section snippets

Motivation

The increasing significance of the topic of adaptive radiotherapy is the result of the presence of challenges that medicine is facing during this kind of treatment.

The literature mentions two main types of challenges in radiotherapy. The first one are inter- and intraobserver delineation variations. There are studies that report that discrepancies between two delineations of the same ROI prepared by two different physicians can be significant [1]. What is more, similar difference can be

The Marching Cubes

The Marching Cubes (MC) algorithm (see [10]) can be used to extract a surface mesh from a segmented voxel grid. However, the Marching Cubes algorithm reproduces the inaccuracies of rough delineations. The resulting surface mesh has significantly noticeable layered structure and is practically unusable for geometrical transformations. An example of this undesirable effect is shown in Fig. 3. Note, that this effect is the most visible after rotating or scaling MC generated meshes and voxelizing

Methodology

To prepare an adequate conformance index a series of experiments has been performed. Each experiment consisted of six steps:

  • 1.

    Extraction of a voxel grid V1 based on a coarse delineation.

  • 2.

    Extraction of a voxel grid V2 based on a careful (reference) delineation.

  • 3.

    Extraction of a surface of the grid V1 as a 3D surface mesh S1 using Marching Cubes algorithm.

  • 4.

    Extraction of a surface of the grid V2 as a 3D surface mesh S2 using Marching Cubes algorithm.

  • 5.

    Poison Reconstruction of the surface S1 based on

Experiments

A sequence of experiments has been prepared to validate our conformance indexes. There were four test datasets: two with prostate delineated twice and two with rectum contours delineated twice. The careful delineations has been converted from voxel grids to surface meshes using only MC algorithm. The other contours have been additionally processed using PSR algorithm. To confirm that the proposed conformance indexes are valid, the corresponding pairs were compared:

  • careful delineations after MC

Conclusions and future work

From the results in Table 1, Table 2 we can draw the following conclusions:

  • Poisson Surface Reconstruction can be used to recover original shapes of organs/tissues. These shapes are more similar to the detailed delineations than the original contour. Although, the improvement does not seem very significant the reconstructed delineations are more suitable for further processing, because of the elimination of local disturbances.

  • Octree depth parameter has high influence on values of the presented

References (12)

There are more references available in the full text version of this article.
View full text