[Paper Review] Trustworthy Multi-phase Liver Tumor Segmentation via Evidence-based Uncertainty

Chuanfei Hu, Tianyi Xia+6 more|arXiv (Cornell University)|May 9, 2023

Spectroscopy Techniques in Biomedical and Chemical Research4 citations

TL;DR

This paper proposes TMPLiTS, a unified framework for trustworthy multi-phase liver tumor segmentation using evidence-based uncertainty estimation via Dempster-Shafer Theory and Dirichlet distribution parameterization. By fusing multi-phase CT images through a theoretically grounded multi-expert mixture scheme, TMPLiTS achieves state-of-the-art performance with high correlation (0.969 and 0.961) to radiologists' annotations and robustness against perturbations, enhancing clinical trustworthiness.

ABSTRACT

Multi-phase liver contrast-enhanced computed tomography (CECT) images convey the complementary multi-phase information for liver tumor segmentation (LiTS), which are crucial to assist the diagnosis of liver cancer clinically. However, the performances of existing multi-phase liver tumor segmentation (MPLiTS)-based methods suffer from redundancy and weak interpretability, % of the fused result, resulting in the implicit unreliability of clinical applications. In this paper, we propose a novel trustworthy multi-phase liver tumor segmentation (TMPLiTS), which is a unified framework jointly conducting segmentation and uncertainty estimation. The trustworthy results could assist the clinicians to make a reliable diagnosis. Specifically, Dempster-Shafer Evidence Theory (DST) is introduced to parameterize the segmentation and uncertainty as evidence following Dirichlet distribution. The reliability of segmentation results among multi-phase CECT images is quantified explicitly. Meanwhile, a multi-expert mixture scheme (MEMS) is proposed to fuse the multi-phase evidences, which can guarantee the effect of fusion procedure based on theoretical analysis. Experimental results demonstrate the superiority of TMPLiTS compared with the state-of-the-art methods. Meanwhile, the robustness of TMPLiTS is verified, where the reliable performance can be guaranteed against the perturbations.

Motivation & Objective

• To address the lack of interpretability and reliability in existing multi-phase liver tumor segmentation methods.
• To jointly perform segmentation and uncertainty estimation in a unified framework for improved clinical trust.
• To model the reliability of segmentation results across multi-phase CECT images using theoretical foundations.
• To ensure robustness against input perturbations while maintaining high segmentation accuracy.
• To provide clinicians with interpretable uncertainty maps highlighting regions of low confidence.

Proposed method

• Evidence-based uncertainty is modeled using Dirichlet distribution to represent segmentation confidence across multi-phase CECT scans.
• Dempster-Shafer Evidence Theory (DST) is applied to formalize segmentation and uncertainty as evidence for fusion.
• A multi-expert mixture scheme (MEMS) fuses evidence from different phases with theoretical guarantees on fusion performance.
• The framework enables independent, interpretable segmentation and uncertainty estimation per phase, followed by reliable fusion.
• The method is designed to be robust against input perturbations, including noise and missing phases.
• Model complexity is balanced to ensure efficiency without sacrificing reliability or accuracy.

Experimental results

Research questions

• RQ1 Can a unified framework jointly achieve accurate segmentation and reliable uncertainty estimation in multi-phase liver tumor segmentation?
• RQ2 How can the reliability of segmentation results across multiple CT phases be explicitly quantified and interpreted?
• RQ3 Can a theoretically grounded fusion mechanism improve the efficiency and reduce redundancy in multi-phase feature fusion?
• RQ4 How does the proposed method perform under input perturbations, such as noise or missing phases?
• RQ5 To what extent does the model's prediction correlate with radiologist-annotated tumor volumes?

Key findings

• TMPLiTS achieves a correlation coefficient of 0.969 between predicted and radiologist-annotated tumor volumes on the internal validation set.
• The method achieves a correlation coefficient of 0.961 on the external validation set, indicating strong agreement with expert annotations.
• The framework demonstrates robustness against three types of perturbations, including noise and missing phases, maintaining reliable performance.
• Uncertainty heatmaps generated by TMPLiTS effectively highlight regions of low confidence, such as incomplete capsules or suspicious patterns, aiding clinical interpretation.
• The multi-expert mixture scheme (MEMS) ensures theoretically sound fusion, reducing redundancy and improving efficiency compared to empirical fusion modules.
• The integration of evidence-based uncertainty via Dirichlet distribution enables explicit, interpretable quantification of model reliability across phases.