Exploring Cycle Consistency Learning in Interactive Volume Segmentation

TL;DR

A simple yet effective cycle consistency loss that regularizes an intermediate segmentation by referencing the accurate segmentation in the starting slice using the same propagation network is proposed.

Abstract

Automatic medical volume segmentation often lacks clinical accuracy, necessitating further refinement. In this work, we interactively approach medical volume segmentation via two decoupled modules: interaction-to-segmentation and segmentation propagation. Given a medical volume, a user first segments a slice (or several slices) via the interaction module and then propagates the segmentation(s) to the remaining slices. The user may repeat this process multiple times until a sufficiently high volume segmentation quality is achieved. However, due to the lack of human correction during propagation, segmentation errors are prone to accumulate in the intermediate slices and may lead to sub-optimal performance. To alleviate this issue, we propose a simple yet effective cycle consistency loss that regularizes an intermediate segmentation by referencing the accurate segmentation in the starting slice. To this end, we introduce a backward segmentation path that propagates the intermediate segmentation back to the starting slice using the same propagation network. With cycle consistency training, the propagation network is better regularized than in standard forward-only training approaches. Evaluation results on challenging AbdomenCT-1K and OAI-ZIB datasets demonstrate the effectiveness of our method.

Figures (4)

• Figure 1: Method overview. Left: our modular framework for interactive volume segmentation. Right: cycle-consistent training for the propagation module, in which all propagation networks share the weights. We introduce a backward segmentation path (shown in red arrows) into a standard training process that consists of only forward segmentation paths (shown in green arrows). "GT" denotes ground truth (shown in green); "seg." denotes model segmentation (shown in red).
• Figure 2: Relative $\mathcal{J\&F}$ improvement on the AbdomenCT-1K dataset. Although the performance drops slightly for 3 organs, we observe improvements on 8 out of the full set of 12 organs, resulting in average improvements across all metrics (Tab. \ref{['tab:main_results_abdomen1k']}).
• Figure 3: Qualitative comparison of our method and baseline on AbdomenCT-1K Ma-2021-AbdomenCT-1K and OAI-ZIB ambellan2019automated. With cycle consistency learning, the STCN model achieves better performance for abdomen vessel segmentation.
• Figure 4: Ablation study on YouTubeVOS 2018 xu2018youtube. $\lambda=0.1$ achieves the best result.