Contour interpolation by deep learning approach

Chenxi Zhao; Ye Duan; Deshan Yang

doi:10.1117/1.JMI.9.6.064003

Contour interpolation by deep learning approach

J Med Imaging (Bellingham). 2022 Nov;9(6):064003. doi: 10.1117/1.JMI.9.6.064003. Epub 2022 Dec 21.

Authors

Chenxi Zhao¹, Ye Duan¹, Deshan Yang²

Affiliations

¹ University of Missouri, Electrical Engineering and Computer Science Department, Columbia, Missouri, United States.
² Duke University, Department of Radiation Oncology, Durham, North Carolina, United States.

Abstract

Purpose: Contour interpolation is an important tool for expediting manual segmentation of anatomical structures. The process allows users to manually contour on discontinuous slices and then automatically fill in the gaps, therefore saving time and efforts. The most used conventional shape-based interpolation (SBI) algorithm, which operates on shape information, often performs suboptimally near the superior and inferior borders of organs and for the gastrointestinal structures. In this study, we present a generic deep learning solution to improve the robustness and accuracy for contour interpolation, especially for these historically difficult cases.

Approach: A generic deep contour interpolation model was developed and trained using 16,796 publicly available cases from 5 different data libraries, covering 15 organs. The network inputs were a $128 \times 128 \times 5$ image patch and the two-dimensional contour masks for the top and bottom slices of the patch. The outputs were the organ masks for the three middle slices. The performance was evaluated on both dice scores and distance-to-agreement (DTA) values.

Results: The deep contour interpolation model achieved a dice score of $0.95 \pm 0.05$ and a mean DTA value of $1.09 \pm 2.30 mm$ , averaged on 3167 testing cases of all 15 organs. In a comparison, the results by the conventional SBI method were $0.94 \pm 0.08$ and $1.50 \pm 3.63 mm$ , respectively. For the difficult cases, the dice score and DTA value were $0.91 \pm 0.09$ and $1.68 \pm 2.28 mm$ by the deep interpolator, compared with $0.86 \pm 0.13$ and $3.43 \pm 5.89 mm$ by SBI. The t-test results confirmed that the performance improvements were statistically significant ( $p < 0.05$ ) for all cases in dice scores and for small organs and difficult cases in DTA values. Ablation studies were also performed.

Conclusions: A deep learning method was developed to enhance the process of contour interpolation. It could be useful for expediting the tasks of manual segmentation of organs and structures in the medical images.

Keywords: contour interpolation; deep learning; medical imaging segmentation.

Abstract

Grants and funding