Assessment of liver metastases radiomic feature reproducibility with deep-learning-based semi-automatic segmentation software.

Background: Good feature reproducibility enhances model reliability. The manual segmentation of gastric cancer with liver metastasis (GCLM) can be time-consuming and unstable.

Purpose: To assess the value of a semi-automatic segmentation tool in improving the reproducibility of the radiomic features of GCLM.

Material And Methods: Patients who underwent dual-source computed tomography were retrospectively reviewed. As an intra-observer analysis, one radiologist segmented metastatic liver lesions manually and semi-automatically twice. Another radiologist re-segmented the lesions once as an inter-observer analysis. A total of 1691 features were extracted. Spearman rank correlation was used for feature reproducibility analysis. The times for manual and semi-automatic segmentation were recorded and analyzed.

Results: Seventy-two patients with 168 lesions were included. Most of the GCLM radiomic features became more reliable with the tool than the manual method. For the intra-observer feature reproducibility analysis of manual and semi-automatic segmentation, the rates of features with good reliability were 45.5% and 62.3% ( < 0.02), respectively; for the inter-observer analysis, the rates were 29.3% and 46.0% ( < 0.05), respectively. For feature types, the semi-automatic method increased reliability in 6/7 types in the intra-observer analysis and 5/7 types in the inter-observer analysis. For image types, the reliability of the square and exponential types was significantly increased. The mean time of semi-automatic segmentation was significantly shorter than that of the manual method ( < 0.05).

Conclusion: The application of semi-automated software increased feature reliability in the intra- and inter-observer analyses. The semi-automatic process took less time than the manual process.