Discerning clinical target volume of endometrial cancer via a lightweight multi-phase delineation framework

TL;DR

This paper introduces a lightweight deep learning framework for accurately delineating clinical target volumes in endometrial cancer radiotherapy using multi-phase CT scans.

Contribution

A novel lightweight framework, NCLNet, is proposed for multi-phase CTV delineation with a new contour-based evaluation metric and lower computational complexity.

Findings

NCLNet achieves a DSC of 0.871 and an ASSD of 0.878 mm with lower computation complexity than nnUNet.

The average modification time by physicians was only 2.9 minutes, with a low modification volume percentage of 3.61%.

The proposed CDSC metric showed higher correlation with clinical modification time than DSC and ASSD.

Abstract

The accurate delineation of the clinical target volume (CTV) is a critical step in precision radiotherapy for endometrial cancer (EC). Multi-phase CT provides more information for delineating the CTV. Our study aims to establish an innovative method for the specific delineation of CTV using multi-phase CT.

Our multi-phase delineation datasets comprise 175 patients who received postoperative pelvic radiotherapy. These datasets include images of contrast-enhanced computed tomography (CECT) and non-contrast-enhanced computed tomography (NECT). Additionally, we introduce a novel framework for automatically segmenting the CTV using a deep learning model. The key component of our framework is the NCLNet, which fuses features from NECT and CECT within a Lightweight Network structure. This structure is optimized using a boundary-aware multi-phase learning strategy that we propose. In addition to the dice similarity coefficient (DSC) and the average symmetric surface distance (ASSD), we propose a novel contour dice similarity coefficient (CDSC) metric to evaluate the accuracy of the predictive outer contour. Three physicians modified the predictive CTV to assess the clinical utility of the proposed method.

The NCLNet achieves a DSC of 0.871 ± 0.027 and an ASSD of 0.878 ± 0.265 mm, with lower computation complexity (5.7 M parameters and FLOPS was 639.1G). In comparison, the widely used nnUNet attains a DSC of 0.860 ± 0.027 and an ASSD of 0.920 ± 0.286 mm, while requiring significantly more parameters (31.0 M) and similar FLOPs (643.8G). The overall evaluation against several benchmarks demonstrates better or comparable performance relative to methods with higher complexity. The average modification time and the modification volume percentage of the automatically delineated CTV were only 2.9 min and 3.61%, respectively. The CDSC of 8 mm thickness was 0.853 ± 0.030, demonstrating higher correlation with the clinical modification time of experts than both DSC and ASSD.

The NCLNet generates high-quality automatic delineation of the CTV for postoperative pelvic radiotherapy in patients with EC.

Figures