A Bi-modal Temporal Segmentation Network for Automated Segmentation of Focal Liver Lesions in Dynamic Contrast-enhanced Ultrasound.

Abstract

Objective: To develop and validate an automated deep learning-based model for focal liver lesion (FLL) segmentation in a dynamic contrast-enhanced ultrasound (CEUS) video.

Methods: In this multi-center and retrospective study, patients with FLL who underwent dynamic CEUS exam were included from September 2021 to December 2021 (model development and internal test sets), and from March 2023 to May 2023 (external test sets). A bi-modal temporal segmentation network (BTS-Net) was developed and its performance was evaluated using Dice score, intersection over union (IoU) and Hausdorff distance, and compared against several segmentation methods. Time-intensity curves (TICs) were obtained automatically from BTS-Net and manually de-lineated by an experienced radiologist, and evaluated by intra-class correlation and Pearson correlation co-efficients. Multiple characteristics were analyzed to evaluate the influencing factors of BTS-Net.

Results: A total of 232 patients (160 men, median age 56 y) with single FLL were enrolled. BTS-Net achieved mean Dice scores of 0.78, 0.74 and 0.80, mean IoUs of 0.67, 0.62 and 0.68, and mean Hausdorff distances of 15.83, 16.01 and 15.04 in the internal test set and two external test sets, respectively. The mean intra-class correlation and Pearson correlation co-efficients of TIC were 0.89, 0.92 and 0.98, and 0.91, 0.93 and 0.99, respectively. BTS-Net demonstrated a significantly higher mean Dice score and IoU in large (0.82, 0.72), homogeneous positive enhanced (0.81, 0.70) or stable (0.81, 0.70) lesions in pooled test sets.

Conclusion: Our study proposed BTS-Net for automated FLL segmentation of dynamic CEUS video, achieving favorable performance in the test sets. Downstream TIC generation based on BTS-Net performed well, demonstrating its potential as an effective segmentation tool in clinical practice.