ESTRO meets Asia 2024 - Abstract Book

S322

Physics – Machine learning models and clinical applications

ESTRO meets Asia 2024

We have released a multi-phase delineation dataset containing CECT, NECT, and CTV images from a cohort of 175 patients who underwent postoperative pelvic radiation therapy. 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, designed to fuse the features of NECT and CECT within a Lightweight Network structure optimized by a proposed boundary-aware multi-phase learning strategy. In addition to the dice similarity coefficient (DSC) and the average symmetric surface distance (ASSD), a novel contour dice similarity coefficient (CDSC) metric was proposed to further evaluate the accuracy of the predictive outer contour. Three radiation oncologists modified the predictive CTV to assess the clinical utility of the proposed method.

Results:

The NCLNet outperforms existing methods delineated the CTV with nearly 28 times lower computation complexity with the DSC, ASSD, and CDSC of 5 mm thickness of the proposed were 0.871±0.027, 0.878±0.265 mm, and 0.795±0.034, respectively The average modification time and volume percentage of the predictive CTV were 2.9 min. The CDSC demonstrating a higher Pearson correlation coefficient with the clinical modification time of physicians.

Conclusion:

The NCLNet ensuring generate high-quality automatic delineation on the CTV of postoperative pelvic radiotherapy for EC.

Keywords: endometrial cancer, automatic delineation