Quantitative analysis and prediction of regional lymph node status in rectal cancer based on computed tomography imaging

Abstract

Objectives

To quantitatively evaluate regional lymph nodes in rectal cancer patients by using an automated, computer-aided approach, and to assess the accuracy of this approach in differentiating benign and malignant lymph nodes.

Methods

Patients (228) with newly diagnosed rectal cancer, confirmed by biopsy, underwent enhanced computed tomography (CT). Patients were assigned to the benign node or malignant node group according to histopathological analysis of node samples. All CT-detected lymph nodes were segmented using the edge detection method, and seven quantitative parameters of each node were measured. To increase the prediction accuracy, a hierarchical model combining the merits of the support and relevance vector machines was proposed to achieve higher performance.

Results

Of the 220 lymph nodes evaluated, 125 were positive and 95 were negative for metastases. Fractal dimension obtained by the Minkowski box-counting approach was higher in malignant nodes than in benign nodes, and there was a significant difference in heterogeneity between metastatic and non-metastatic lymph nodes. The overall performance of the proposed model is shown to have accuracy as high as 88% using morphological characterisation of lymph nodes.

Conclusions

Computer-aided quantitative analysis can improve the prediction of node status in rectal cancer.

Appendix

This appendix has been provided by the authors to give readers additional information about their work.

Quantitative measurements of lymph nodes

To achieve comprehensive characterization of a lymph node, 19 parameters were computed automatically based on segmented image. Among which seven were chosen to consist of a compact but highly representative subset. The detailed definitions of the parameters are listed below:

• Area: the area of the node;

• Major axis length & minor axis length: the major & minor axis of the ellipse that covers the node with same normalized second central moments;

• Solidity: the ratio of the area of the node and its convex area;

• Density: the ratio between the summation of grey value within the node and its area;

• Heterogeneity: Fraction of pixels that deviate more than a certain range (10% default) from the average intensity.

• Fractal dimension: Minkowski dimension of the boundary of the node, computed by box-counting method.

• Minkowski dimension: a way of determining the fractal dimension of set S in a Euclidean space Rⁿ. It is estimated by limit of ${\lim_{{\tau \to 0}}}\frac{{{N_{\tau }}}}{{1/\varepsilon }}$, where N(ε) is the number of boxes of side length ε required to cover the set S.